Chaos, Solitons and Fractals 209 (2026) 118383

Contents lists available at ScienceDirect

Chaos, Solitons and Fractals

journal homepage: [Link]/locate/chaos

Gudermannian-based error-shaped PID controller tuned via THRO

for nonlinear CSTR temperature control
Serdar Ekinci a , Gokhan Yuksek b, Ridvan Firat Cinar c , Cebrail Turkeri c ,
Dacheng Li d,*
a
Department of Computer Engineering, Bitlis Eren University, Bitlis, 13100, Turkey
b
Department of Electrical-Electronics Engineering, Batman University, Batman, 72100, Turkey
c
Department of Computer Engineering, Batman University, Batman, 72100, Turkey
d
Birmingham Centre for Energy Storage, School of Chemical Engineering, University of Birmingham, Edgbaston, B15 2TT, UK

A R T I C L E I N F O A B S T R A C T

Keywords: Temperature regulation in nonlinear continuous stirred-tank reactors (CSTRs) is challenging due
Gudermannian function-based PID controller to the strong coupling between reaction kinetics and heat transfer, which leads to state-dependent
Temperature regulation of CSTR process gain variations and asymmetric transient behaviour. This study addresses the limitation of clas­
Tianji's horse racing optimization
sical PID controllers under such nonlinear conditions by modifying the error-processing stage
Nonlinear system
Robustness
rather than the controller structure itself. A Gudermannian-based error-shaping mechanism is
introduced into a standard PID architecture to obtain a smooth, bounded, and continuously
differentiable transformation of the tracking error. This mapping attenuates large error magni­
tudes while preserving sensitivity near the setpoint, resulting in an error-dependent modulation
of the proportional, integral, and derivative actions. The controller parameters are tuned using
Tianji's Horse Racing Optimizer (THRO) within a closed-loop nonlinear simulation framework.
Proposed gd-PID controller is evaluated on a nonlinear CSTR temperature model and compared
with PI, PIDf, two-degree-of-freedom PID, and fractional-order PID controllers under identical
conditions. Simulation results show that the gd-PID achieves a rise time of 0.586 min and a
settling time of 1.852 min, while limiting overshoot to 0.2186%. The integral absolute error is
reduced to 2.858, representing a reduction of more than 55% compared with conventional PID-
based designs. Additional tests involving multi-step reference changes, feed-temperature distur­
bances, and measurement noise confirm that the proposed controller maintains stable and
physically consistent behaviour across wide operating regions. The results indicate that
Gudermannian-based error shaping, combined with THRO-based tuning, provides an effective
and computationally simple approach for nonlinear CSTR temperature control.

1. Introduction

Reactor temperature control becomes difficult when heat release directly influences reaction rates, leading to nonlinear feedback
within the process dynamics. Such behaviour affects operational stability, energy efficiency and product consistency in reactive

* Corresponding author: School of Chemical Engineering, University of Birmingham, Birmingham B15 2TT, UK.
E-mail addresses: sekinci@[Link] (S. Ekinci), [Link]@[Link] (G. Yuksek), [Link]@[Link] (R.F. Cinar),
[Link]@[Link] (C. Turkeri), [Link].6@[Link] (D. Li).

[Link]
Received 16 December 2025; Received in revised form 30 March 2026; Accepted 18 April 2026
Available online 29 April 2026
0960-0779/© 2026 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
([Link]
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

systems, making temperature regulation a central issue in reaction engineering. For continuous stirred-tank reactors (CSTRs), this
challenge is more pronounced, since heat generation and heat removal must remain closely balanced. Minor temperature deviations
may therefore modify reaction rates, shift equilibrium conditions, or induce oscillatory and unstable operating regimes [1–3]. This
sensitivity arises from the Arrhenius-type dependence of reaction rates on temperature and the enthalpy accumulation that reinforces
deviations through nonlinear feedback. Consequently, CSTR dynamics exhibit asymmetric responses, state-dependent gain variations
and substantial changes in transient characteristics across different operating conditions. Under these conditions, reaching stable and
accurate temperature regulation demands control strategies capable of adapting to nonlinearities while preventing excessive
corrective action and preserving smooth actuator behaviour [4,5]. Foundational studies in nonlinear process control highlight that
CSTRs exhibit multiple steady states, input–output gain reversals and enthalpy-driven bifurcations, making them a testbed for eval­
uating advanced control architectures [6,7].
PID controllers remain widely used in industrial settings because of their simple structure, transparent tuning logic and low
computational cost [8–10]. Here, standard PID tuning methods assume approximately linear relationship between the manipulated
variable and the controlled variable. This assumption holds only in narrow operating regions and deteriorates when internal reaction
kinetics dominate the energy balance. The proportional, integral and derivative actions interact in a coupled manner, and their
combined effect varies with temperature-dependent gain changes. This interaction makes it difficult to balance quick response, low
overshoot and disturbance attenuation simultaneously. Furthermore, a single-loop structure forces the same control path for tracking
and disturbance rejection, limiting flexibility when the reactor experiences abrupt heat-release variations or strongly asymmetric
cooling behaviour. These characteristics reduce the reliability of classical PID controllers in nonlinear CSTR environments [11–13].
To address these limitations, several extensions of PID control have been proposed. Fractional-order PID (FOPID) [14,15] in­
troduces additional integrative and derivative orders, offering finer shaping of the closed-loop frequency response. Two-degree-of-
freedom PID (2DOF-PID) systems [16–18] enable independent weighting for reference tracking and disturbance rejection. Filtered-
derivative PID (PIDf) [19,20] improves noise attenuation by modifying the derivative term. Although these approaches increase
tuning flexibility, they also expand the parameter space and complicate gain interactions. Under nonlinear dynamics, performance
becomes highly sensitive to parameter selection, and deterministic tuning rules often fail to deliver stable results across different
operating regions. While various nonlinear control strategies such as fuzzy logic [21–23], sliding-mode designs [24,25], PIDA
structures [26,27] and cascaded controllers [28–30] have been explored in related applications, PID-based architectures remain the
dominant choice in industry due to their simplicity, transparency and ease of deployment. The tuning and searching problems thus
advances into a nonlinear and multidimensional search task, motivating the use of optimization-based methods [31,32].
A wide range of optimization algorithms has been adopted for PID tuning. Math-based approaches such as Ziegler–Nichols [33–35]
rules, pole-placement strategies, convex formulations and nonlinear programming [36] provide analytical tuning guidelines but may
lose validity when the system deviates from local linearity. Nature-inspired metaheuristics introduce alternative search mechanisms
capable of exploring non-convex tuning landscapes. Evolutionary techniques such as genetic algorithm (GA) [37,38] and differential
evolution (DE) [39], swarm-based algorithms such as particle swarm optimization (PSO), firefly algorithm (FA) [40], grey wolf
optimizer (GWO) [41], whale optimization algorithm (WOA) [42] and cuckoo search [43], and physics-driven methods such as
simulated annealing [44], gravitational search algorithm (GSA) [45], electromagnetism-like optimization [46] and atom search
optimization (ASO) [47,48] have all been applied to PID tuning. Human-behaviour-inspired approaches, including teaching–learning
based optimization [49], physics based algorithm [50], harmony search [51] and imperialist competitive algorithm [52], extend this
ecosystem by providing different population-update dynamics. These methods improve search diversity but often demand algorithm-
specific hyperparameters and may generate inconsistent gain sets under nonlinear system behaviour.
Although a wide range of metaheuristic optimizers has been applied to nonlinear process control problems, their use in CSTR
temperature regulation is often limited to parameter tuning rather than direct handling of process nonlinearity. In most cases, the
optimizer operates externally to the control law and does not alter the underlying error–control relationship governing the reactor
dynamics. As a result, the strong temperature–reaction coupling and state-dependent gain characteristics of the CSTR remain primarily
addressed through controller structure rather than through the optimization mechanism itself. This motivates the exploration of
complementary approaches that modify the control behaviour at the error-processing level, rather than relying solely on increasingly
complex optimization strategies.
Recent studies on CSTR temperature control reflect a raising attention in optimization-assisted PID frameworks. GA-based tuners
[53] are shown to improve rise times and overshoot suppression for exothermic reactions, while hybrid Firefly–Sparrow controllers
provide better settling behaviour under nonlinear and time-delayed conditions. Swarm-based strategies such as PSO [54,55], ant
colony optimization (ACO) [56], coot bird optimization algorithm (CBOA) [57] and artificial bee colony (ABC) [58,59], firefly al­
gorithm (FA) [60] have also been applied to CSTR temperature dynamics, producing faster convergence but sometimes exhibiting
sensitivity to initial population distribution. More recent metaheuristics, including grey wolf optimizer (GWO) [61], cuckoo search
[62], bat algorithm [63,64], and Harris hawk optimization (HHO) [65], demonstrate improved exploration capability but may
generate inconsistent gain sets when the reactor undergoes rapid thermal shifts. Hybrid formulations combining GA–PSO [66],
PSO–DE [67], PSO-AIW [68] or FA–GWO [69] have been proposed to balance global and local search properties, while HGSO-based
tuning strategies [70] aim to stabilize performance across a wider operating range. Beyond the metaheuristics, advanced model-based
approaches such as nonlinear MPC–PID cascades [71], constraint-guided PID tuning and state-dependent Riccati equation adjustments
[72] have been employed to achieve smoother transitions. Intelligent techniques, including reinforcement-learning-driven PID tuning
[73], ANFIS-based adaptive layers and neural-network-assisted prediction mechanisms, further extend the design space [74].
Although these methods provide meaningful improvements, many of them rely on discontinuous nonlinear mappings, auxiliary
models, broad tuning landscapes, or algorithm-specific hyperparameters [75]. As a result, they may not fully capture the heat-release-

2
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

driven nonlinearities, enthalpy feedback and state-dependent sensitivity that characterize CSTR behaviour, leaving room for controller
structures that directly shape the error signal and maintain smooth control behaviour across large thermal variations. But none of these
approaches integrates a smooth bounded nonlinear mapping into the PID structure while simultaneously tuning all controller pa­
rameters through a stable global-search mechanism, leaving a clear methodological gap for nonlinear CSTR regulation.
While these approaches improve performance under certain conditions, they do not adequately address how the control action
evolves when large temperature mismatches occur. In nonlinear CSTR operation, rapid variations in heat generation can amplify small
modelling errors, particularly when the controller responds through fixed or non-smooth error-processing mechanisms. As a result,
many tuning strategies focus on optimizing tracking or transient performance indices without explicitly regulating how the error signal
influences the control action across different operating regions. This limitation may lead to overly aggressive responses far from the
setpoint, while insufficient sensitivity is maintained near equilibrium conditions. These observations indicate that modifying the error-
processing mechanism itself, rather than further extending controller structure or tuning complexity, can provide a more direct way to
regulate control behaviour under nonlinear conditions [76,77]. From this perspective, transformation of the error signal emerges as an
alternative mechanism for shaping closed-loop behaviour. However, this aspect has received limited attention in nonlinear chemical
process control, where most existing approaches rely on structural extensions such as fractional-order formulations, gain scheduling,
or multi-degree-of-freedom designs. Such methods often increase model complexity or introduce additional parameters without
directly addressing how the control action is modulated across varying operating conditions. In contrast, a smooth and bounded error
transformation can enable continuous adaptation of the control response while preserving stability and avoiding abrupt changes in the
control signal.
Motivated by this gap, the present study develops a PID architecture in which a single smooth and bounded analytic mapping is
embedded into the error-processing stage of a classical controller. Specifically, the Gudermannian function [78,79] is integrated into
the error path, resulting in the gd-PID controller. This mapping attenuates large error magnitudes while maintaining sensitivity near
the reference point, thereby introducing a natural gain-modulation effect across the proportional, integral, and derivative components.
In this way, the control action is reshaped under large temperature deviations without altering the fundamental PID structure or
introducing additional degrees of freedom.
The controller parameters are determined using Tianji's horse racing optimization (THRO) [80], which provides an efficient global
search capability through rank-based learning and adaptive stochastic updates. By combining a smooth nonlinear error transformation
with an effective optimization strategy, the proposed approach aims to achieve fast, stable, and consistent temperature regulation
under disturbances, measurement noise, and varying operating conditions. Accordingly, the primary objective of this study is to
develop a smooth and bounded error-shaping mechanism within a classical PID framework and to evaluate its effectiveness in
nonlinear CSTR temperature control.
The main contributions of this study are summarized as follows:

• A Gudermannian-based error-shaping mechanism is analytically embedded into a classical PID controller, introducing a smooth
and bounded nonlinear transformation without altering the original PID structure.
• Unlike saturation blocks, gain scheduling, or fractional extensions, the proposed approach modifies the control behaviour solely
through continuous error transformation, avoiding discontinuities and parameter proliferation.

Fig. 1. Continuous stirred tank reactor with cooling jacket.

3
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

• Tianji's Horse Racing Optimizer is applied to jointly tune the linear PID gains and nonlinear shaping parameters, representing one
of the first uses of THRO in nonlinear chemical process control.
• A comparison with PI, PIDf, 2DOF PID, and FOPID controllers under identical nonlinear CSTR conditions demonstrates consistent
improvements in transient speed and cumulative error.
• Extensive simulations under reference changes, feed-temperature disturbances, and measurement noise confirm that the proposed
structure maintains stable and physically consistent behaviour across wide operating regions.

The rest of this paper is organized as follows. Section 2 presents the nonlinear CSTR model and the fundamental equations that
describe its temperature and concentration dynamics. Section 3 details the principles of the Tianji's Horse Racing Optimization al­
gorithm, which is used to tune the controller parameters. Section 4 introduces the Gudermannian-function-based PID frame and details
its mathematical properties. Section 5 defines the optimization problem and the performance index used for evaluating the controller
behaviour. Section 6 reports comparative simulation results against several benchmark controllers and optimization methods. Section
7 examines the robustness of the proposed approach under reference changes, disturbances and measurement noise. Finally, Section 8
summarizes the main findings and discusses potential directions for future research.

2. Continuous stirred tank reactor model

Continuous stirred-tank reactors (CSTR) constitute one of the canonical platforms for studying the interaction between chemical
reaction kinetics and thermal management in liquid phase systems. A schematic diagram of the continuous stirred-tank reactor
considered in this study is shown in in Fig. 1. The reactor is supplied by a continuous feed stream entering from the top, characterized
by a fixed inlet concentration and inlet temperature. An equal volumetric outflow leaves the reactor through the outlet line, ensuring
constant liquid volume throughout operation. Inside the vessel, an agitator maintains complete mixing, allowing the reactant con­
centration CA and reactor temperature T to be modelled as spatially uniform state variables. The exothermic reaction A→B takes place
within the well-mixed liquid phase, generating heat that directly couples concentration and temperature dynamics. Thermal regu­
lation is achieved through an external cooling jacket surrounding the reactor wall. The jacket temperature Tj is the manipulated input
of the control system and governs the rate of heat removal from the reactor. Variations in Tj influence the internal energy balance
without directly altering the mass flow structure, making jacket temperature the primary control variable for regulating the nonlinear
reactor temperature dynamics.
The reaction considered in this work is an exothermic, irreversible, first-order conversion of species A into product B. Since the
reaction releases heat at a temperature-dependent rate, the complete reactor behaviour results from a delicate balance between heat
generation and heat extraction. Even small deviations between these contributions may modify the qualitative nature of the steady-
state behaviour. It is well established that for sufficiently exothermic reactions, this balance can produce multiple admissible steady
states along with regions of open-loop thermal instability, making the CSTR a widely used benchmark for nonlinear control analysis.
To formulate the dynamic model, the reactor is assumed to operate at a constant liquid volume V, with constant physical properties
such as density ρ and heat capacity cp . Since the jacket temperature is directly controlled, no separate dynamics are assigned to the
jacket; instead, heat transfer is modelled through an overall coefficient UA, representing the combined effects of heat-transfer area and
thermal resistance. Under these assumptions, the concentration CA of reactant A evolves according to the transient mass balance
dCA ( )
V = F CAf − CA − rV, (1)
dt

where F denotes the volumetric flow rate, CAf is the inlet concentration, and r is the reaction rate. For an irreversible first-order re­
action, the rate follows the Arrhenius expression
( )
ΔE
r = k0 exp − CA , (2)
RT

where k0 is the pre-exponential constant, ΔE is the activation energy, R is the universal gas constant, and T is the reactor temperature.
This expression highlights the exponential dependence of the reaction rate on temperature, indicating that even small variations in
reactor temperature can lead to significant changes in reaction kinetics. This temperature sensitivity is a key source of nonlinearity in
the CSTR dynamics and plays a central role in shaping the coupled behaviour of concentration and temperature states. Combining Eqs.
(1) and (2), the concentration dynamics take the nonlinear coupled form
( )
dCA F ( ) ΔE
= CAf − CA − k0 exp − CA . (3)
dt V RT
The temperature dynamics are governed by the total energy balance, which accounts for convective energy transport due to inlet/
outlet flow, reaction-generated heat, and heat loss to the jacket:
dT ( ) ( )
V ρcp = Fρcp Tf − T + ( − ΔH) rV − UA T − Tj , (4)
dt

where Tf denotes the inlet temperature and ΔH is the (negative) heat of reaction. Substituting Eq. (2) into Eq. (4) gives the explicit

4
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

nonlinear temperature equation;

( )
dT F ( ) − ΔH ΔE UA ( )
= Tf − T + k0 exp − CA − T − Tj . (5)
dt V ρcp RT V ρcp

Eqs. (3) and (5) constitute the complete dynamic model of the CSTR. Their strong coupling arises from the exponential dependence
of the reaction rate on temperature; consequently, the reactor temperature directly affects the rate at which reactant is consumed,
while the concentration determines the amount of heat released. This reciprocal interaction is the root cause of nonlinearities such as
multiplicity of steady states and thermal runaway.
At steady state, the time derivatives vanish and the algebraic relations become;
( )
F( ) ΔE
0= CAf − CA − k0 exp − CA , (6)
V RT
( )
F( ) − ΔH ΔE UA ( )
0= Tf − T + k0 exp − CA − T − Tj . (7)
V ρcp RT V ρcp

Using the parameter set reported in [7], the simultaneous solution of Eqs. (6)–(7) yields the nominal operating point
Tj = 280 K, T = 304.167553089807 K and CA = 0.977403565332 mol/L.
This equilibrium is taken as the reference condition for evaluating the Gudermannian-based PID control strategy developed in this
work.
The mathematical model of the CSTR considered in this study is developed under the following assumptions:

• The reactor operates at constant volume with equal inlet and outlet flow rates.
• The reactor contents are perfectly mixed, ensuring uniform temperature and concentration throughout the vessel.
• The physical properties such as density and heat capacity are assumed to be constant.
• The reaction follows an irreversible first-order kinetics governed by the Arrhenius law.
• The heat transfer between the reactor and the cooling jacket is represented by an overall heat transfer coefficient.
• The dynamics of the cooling jacket are neglected, and the jacket temperature is treated as the control input.

These assumptions are widely adopted in nonlinear CSTR modelling and provide a tractable yet representative framework for
analyzing temperature control dynamics.

3. Tianji's horse racing optimization

Tianji's Horse Racing Optimization (THRO) [80] operates with two equal-sized populations of d-dimensional solutions, referred to
as Tianji's horses and the King's horses. After initialization, both sets are evaluated and sorted in ascending order of fitness so that index
i always corresponds to horses of comparable rank. Optimization proceeds by pairing the i-th horses of both sides and updating them
according to which horse wins the race. To preserve the algorithm's strategic nature while maintaining mathematical clarity, all update
rules can be expressed through a combination of a weighted learning operator, random perturbation, and an occasional Lévy-driven
exploration term. The weight factor controlling the influence between two horses is defined as p = 1 − Tt , while the directional learning
operator is written as D(a, b) = p a + (1 − p) b. THRO relies on two disturbance components: a random perturbation R = β ⋅ rand( −
1, 1) and a Lévy perturbation L = γ ⋅ Levy(λ), which together allow both short and long-range updates as required throughout the
competition phases.
The first scenario occurs when Tianji's slowest horse is actually faster than the King's slowest horse, meaning f(xTi ) < f(xKi ) at the
lowest rank levels. Tianji wins this race, and to maintain this advantage, the algorithm strengthens Tianji's slowest horse by pulling it
toward Tianji's fastest horse while introducing a small stochastic correction. A compact form of this update is expressed as:
( )
xʹTi = D xTi , xTf + R, (8)

where xTf denotes Tianji's fastest horse. At the same time, the King's slowest horse attempts to catch up by adopting a guided movement
toward the updated Tianji horse, given by:

xʹKi = D(xKi , xTi ) + R. (9)

In the second scenario, Tianji's slowest horse is slower than the King's slowest horse, i.e., f(xTi ) > f(xKi ). Tianji intentionally sac­
rifices this weakest horse against the King's fastest horse, mimicking the original story's strategic loss. Because Tianji expects to lose, his
slowest horse learns not from the King but from a randomly selected horse within his own population so that exploration is increased.
This update is written as:

xʹTi = D(xTi , xTr ) + R, (10)

where xTr is a random horse from Tianji's set. Meanwhile, the King strengthens his fastest horse by updating it toward the best horse on
his side:

5
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

( )
xʹKi = D xKi , xKf + R. (11)

The third scenario happens when the slowest horses of both sides run at nearly the same speed while Tianji's fastest horse is superior
to King's fastest horse. Under this condition, Tianji uses his fastest horse to challenge the King's fastest horse and wins. To preserve
leadership, Tianji's fastest horse is reinforced through a focused update toward Tianji's own best horse:
( )
xʹTi = D xTi , xTf + R. (12)

Since the King loses this race at the top tier, his fastest horse attempts to match Tianji's by adopting,
( )
xʹKi = D xKi , xTf + R. (13)

The fourth scenario is triggered when the slowest horses of both sides have similar speed but Tianji's fastest horse is slower than the
King's fastest horse. Following Tianji's historical strategy, the algorithm deliberately sends Tianji's slowest horse to compete against the
King's strongest horse, accepting an intentional loss. Because Tianji expects no advantage here, this slowest horse is driven toward a
randomly selected Tianji horse with an exploration-promoting step:

xʹTi = D(xTi , xTr ) + R. (14)

The King, in contrast, focuses on supporting his fastest horse by updating it toward the King's fastest reference horse:
( )
xʹKi = D xKi , xKf + R. (15)

The fifth scenario arises when both the slowest and the fastest horses of the two populations are nearly equal in performance. In this
fully balanced case, Tianji again sends his slowest horse against the King's fastest horse, which produces another structured loss.
Because neither side has a clear advantage, both updates combine directional movement with exploratory Lévy motion. Tianji's horse
is updated as:

xʹTi = D(xTi , xTr ) + L, (16)

and the King's horse also adopts a mixture of learning and perturbation,
( )
xʹKi = D xKi , xKf + R. (17)

After all i-indexed pairs are processed, THRO enters a refinement stage in which candidate solutions are locally adjusted around
high-ranking agents. This step emphasizes exploitation by reducing search dispersion and guiding updates toward regions associated
with lower objective values, thereby supporting convergence in later iterations. In this refinement, Tianji's updated horses move to­
ward the global best using:
( )
xʹʹTi = xʹTi + α xbest − xʹTi + L, (18)

while King's horses adjust toward the refined Tianji horses through:
( )
xʹʹKi = xʹKi + α xʹTi − xʹKi + R. (19)

After the refinement step, the updated population replaces the previous one and fitness values are recalculated before proceeding to
the next iteration. The global best solution is revised accordingly, and this cycle continues until the stopping criterion is met. Through
its five competitive scenarios and the two-stage update mechanism, THRO gradually balances exploratory and exploitative behaviour.
While strategy-driven races guide the population toward promising regions, stochastic updates and Lévy-based perturbations help
preserve diversity and allow occasional long-range moves. This combination makes it less likely for the search process to remain
trapped in local optima, even when the objective landscape contains multiple competing minima.

Pseudo-Code for THRO

Set the number of horses nfor two populations and the number of iterations T.
Randomly initialize Tianji's horses {xT1 , …, xTn } and King's horses {xK1 , …, xKn }.
Evaluate their fitness f(xTi ) and f(xKi ).
Sort both populations in ascending order.
While the stopping condition is not satisfied do
For each pairing i do
If f(xTi ) < f(xKi ) then
Update xTi based on Eq. (8);
Update xKi based on Eq. (9);
Evaluate their fitnesses;
ElseIf f(xTi ) > f(xKi ) then
Update xTi based on Eq. (10);
Update xKi based on Eq. (11);
Evaluate their fitnesses;
( ) ( )
ElseIf f xTf < f xKf then
Update xTi based on Eq. (12);
(continued on next page)

6
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

(continued )
Update xKi based on Eq. (13);
Evaluate their fitnesses;
( ) ( )
ElseIf f xTf > f xKf then
Update xTi based on Eq. (14);
Update xKi based on Eq. (15);
Evaluate their fitnesses;
Else
Update xTi based on Eq. (16);
Update xKi based on Eq. (17);
Evaluate their fitnesses;
End If
End For
For each horse i do
Update xTi based on Eq. (18);
Update xKi based on Eq. (19);
(continued on next page)

Fig. 2. Pseudocode of the Tianji's Horse Racing Optimization.

7
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

(continued )
Evaluate their fitnesses;
End For
End While
Return the best horse.

The computational complexity of the proposed approach stems from two components: the gd-PID control law and the THRO-based
parameter tuning procedure. The gd-PID preserves the constant-time complexity of a classical PID controller, i.e., O (1), since the
Gudermannian mapping is evaluated in closed form at each sampling instant using a fixed number of elementary operations. Although
this mapping involves transcendental functions, it does not introduce any iterative computation, memory growth, or sample-
dependent scaling, and therefore does not alter the asymptotic computational order of the control law. The dominant computa­
tional cost is associated with the THRO optimizer used during the offline tuning stage. For a population size Np , number of iterations Ni ,
( )
and simulation horizon length Nt , the overall tuning complexity scales as O Np Ni Nt , consistent with population-based metaheuristic
optimization methods. Since THRO is employed exclusively for offline parameter optimization, the online execution cost of the
proposed controller remains equivalent to that of a standard PID controller (Fig. 2).

4. Gudermannian function and application to PID

4.1. The mathematical fundamentals of the Gudermannian function

The Gudermannian function, denoted as gd(x), establishes a smooth and strictly increasing mapping between hyperbolic and
circular angle measures without the use of complex numbers. This property was first formalized in the eighteenth and nineteenth
centuries through the works of Lambert and Gudermann, and it has since appeared in a variety of mathematical contexts, including
geodesy, conformal mapping, and special function theory. Its analytical foundation reveals a direct relationship between hyperbolic
and trigonometric structures, which is visually represented in Fig. 3.
The most common definition of the Gudermannian function is given as:
gd(x) = arctan(sinhx). (20)
Here, the hyperbolic sine is mapped to a circular angle through the arctangent transformation. Several equivalent closed-form
identities arise from the interplay between circular and hyperbolic functions:
( x)
gd(x) = arcsin(tanhx), gd(x) = 2arctan tanh , (21)
2

each of which highlights the function's ability to convert unbounded hyperbolic inputs into a bounded angular range. This mapping is
further illustrated in Fig. 3, where a common stereographic projection links the areas of the circular and hyperbolic sectors. Through
this geometric relationship, the three variables:

Fig. 3. Geometric relation linking the hyperbolic and trigonometric representations of the Gudermannian function.

8
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

( ) (ψ )
ϕ
ψ = gd− 1 (ϕ), ϕ = gd(ψ ), s = tan = tanh , (22)
2 2

are shown to represent hyperbolic and circular measures of the same geometric configuration.
A fundamental characterization of the Gudermannian function is given by its integral representation shown below:
∫x
gd(x) = sech(u) du, (23)
0

where the integrand is the hyperbolic secant. This form makes obvious that the derivative of the function is:
d 1
gd(x) = sech(x) = . (24)
dx cosh(x)
The hyperbolic secant produces high sensitivity near the origin and rapidly decaying influence as ∣ x ∣ increases. Consequently,
gd(x) behaves almost linearly for small arguments but saturates smoothly to the finite limits:
π π
− < gd(x) < . (25)
2 2
The Gudermannian function is a well-known analytic mapping that connects hyperbolic and trigonometric expressions, and its
properties have been discussed in various mathematical sources. Beyond its formal definition, the function is characterized by a
smooth transition and a bounded response, properties that distinguish it from hard saturation or piecewise nonlinear mappings. Such
features make it suitable for control-oriented formulations in which nonlinear growth needs to be moderated without introducing
abrupt changes.
The saturating S-shaped curve shown in Fig. 4 is symmetric, continuously differentiable, and free of sharp transitions. These
properties have historically motivated the use of the Gudermannian function in applications such as the Mercator projection, where
latitude must be mapped continuously and monotonically while avoiding singular behaviour near extreme values. Although smooth
error-shaping functions such as sigmoid, hyperbolic tangent, or arctangent have been previously employed in nonlinear PID designs,
the Gudermannian function exhibits several distinctive characteristics that motivate its use in this study. In comparison with
commonly used smooth nonlinear mappings such as sigmoid, hyperbolic tangent, and arctangent functions, the Gudermannian
function offers an alternative smooth and bounded transformation characterized by its intrinsic relationship between hyperbolic and
trigonometric functions. While these alternative mappings exhibit similar saturation behaviour, their shaping characteristics are
typically adjusted through scaling parameters. In contrast, the Gudermannian mapping possesses a fixed analytic structure with a well-
defined derivative profile, which naturally supports smooth and continuous modulation of the effective control gain.
A key feature of the Gudermannian mapping lies in the behaviour of its derivative, which decays smoothly and symmetrically as the
error magnitude increases. This results in a continuous modulation of the effective feedback gain: high local sensitivity is preserved
near the operating point, while control aggressiveness is progressively attenuated for large deviations. Importantly, this gain modu­
lation emerges from the analytic structure of the function itself rather than from imposed saturation limits.
For nonlinear processes such as CSTRs, where the system gain varies significantly with temperature due to Arrhenius-type reaction
kinetics, this analytic gain modulation provides a natural mechanism for matching controller sensitivity to the intrinsic nonlinear
behaviour of the plant. This structural distinction differentiates the proposed gd-PID formulation from conventional smooth
saturation-based nonlinear PID approaches.

Fig. 4. Input–output characteristic of the Gudermannian function employed for proposed gd-PID controller.

9
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

4.2. Novel Gudermannian function-based PID (gd-PID) controller

The nonlinear behaviour of the Gudermannian function, illustrated in Fig. 4, provides a smooth saturating transformation that can
be advantageously embedded into conventional PID structures to improve transient performance and enhance robustness against large
tracking errors. Motivated by its strictly monotonic, bounded, and continuously differentiable form, a new control architecture,
referred to as the Gudermannian-function-based PID (gd-PID) controller, is developed in this study. The complete block diagram of the
proposed structure is presented in Fig. 5.
In the gd-PID controller, the tracking error e(t) is first scaled through a linear gain G1 to obtain:
k(t) = G1 e(t). (26)
This transitional signal is then processed by the nonlinear activation stage, where the Gudermannian function appears in its
hyperbolic-sine representation:
l(t) = G2 gd(k(t) ) = G2 arctan(sinh(k(t) ) ). (27)
The graph of the transformation in Fig. 4 indicates that the function arctan(sinh(x) ) varies in an almost linear manner around the
origin, before gradually flattening as the magnitude |x| increases. As a result, small error values lead to relatively strong corrective
responses, whereas larger deviations are moderated in a more progressive way. This transition helps limit abrupt changes and
excessive high-frequency activity, without relying on hard saturation blocks or piecewise limiting functions.
The output of the nonlinear stage, l(t), serves as the sole input to the proportional, integral, and derivative channels of the
controller:
∫t
Ns
u(t) = Kp l(t) + Ki l(τ) dτ + Kd l(t) (28)
0 s+N
The derivative part incorporates a first-order filter of bandwidth N, preventing noise amplification in high-frequency regions.
Because the same nonlinear signal l(t) enters all three branches, the effective proportional, integral, and derivative actions become
error dependent. This error-dependent modulation creates a natural gain-scheduling effect governed by the derivative of the
Gudermannian function:
d 1
gd(k) = sech(k) = . (29)
dk cosh(k)
Since sech(k) decreases rapidly for large |k|, the equivalent gains of the controller evolve as:

Keff
p (t) = Kp sech(k(t) ), Ki (t) = Ki sech(k(t) ), Kd (t) = Kd sech(k(t) ).
eff eff
(30)

This modulation allows the controller to react strongly when the error remains close to the reference, while limiting excessive
control effort during large disturbances or early transients. In such cases, overshoot is reduced and actuator activity remains bounded,
whereas the steady-state accuracy associated with classical PID control is largely maintained. From a dynamic analysis viewpoint, this
behaviour can be interpreted through the local slope of the Gudermannian function, which directly governs the equivalent loop gain of
the controller. Near the operating point, the derivative of the Gudermannian mapping remains high, leading to an increased effective
proportional and integral action and thereby accelerating the initial transient. As the error magnitude grows, the decreasing slope
progressively reduces the effective gain, which suppresses aggressive control action and improves damping during large excursions.
As illustrated in Fig. 5, the gd-PID architecture alters only the error-processing stage, leaving the remaining PID structure un­
changed. This design choice maintains compatibility with conventional tuning practices and introduces a controlled nonlinear degree
of freedom through the parameters G1 , G2 , and the geometric properties of the Gudermannian mapping. Unlike saturation blocks or
piecewise-linear gain scheduling, the nonlinear shaping here arises from a single analytic expression, avoiding discontinuities in the
control signal.
The overall behaviour of the gd-PID controller is therefore governed by the combined effect of the Gudermannian transformation,
the underlying PID dynamics, and the derivative-based attenuation mechanism. Together, these elements influence stability margins
and noise sensitivity, while improving transient responses in operating regimes where large errors may otherwise trigger actuator
saturation or oscillatory behaviour. The following section evaluates these effects in closed-loop simulations and compares the gd-PID

Fig. 5. Block diagram of Gudermannian function-based PID (gd-PID) controller.

10
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

response with that of conventional PID configurations.

5. Objective function and definition of nonlinear optimization problem for CSTR process

Nonlinear tuning of the proposed gd-PID controller is carried out using the continuous stirred tank reactor (CSTR) model illustrated
in Fig. 6. In this configuration, the controller manipulates the jacket temperature in response to the reference signal in order to regulate
the reactor temperature around a selected operating point. The parameter adjustment process is governed by the THRO algorithm,
which operates in an outer optimization loop and updates the controller gains based on performance measures extracted from suc­
cessive closed-loop CSTR simulations.
To introduce a noticeable transient into the reactor dynamics, the reference temperature is stepped from 304.16755 K to
314.16755 K at t = 1 min, which corresponds to a 10 K change. This operating condition imposes a sudden thermal load on the system
and is deliberately selected to evaluate the controller's ability to handle abrupt temperature variations while maintaining closed-loop
stability and acceptable steady-state accuracy.
Fig. 6 presents a high-level block-diagram representation of the proposed control architecture, illustrating the signal flow between
the reference input, the gd-PID controller, the nonlinear CSTR plant, and the THRO-based outer-loop tuning mechanism. This sche­
matic provides a concise visual summary of the main components and their interconnections prior to the presentation of detailed
simulation and benchmarking results. For each candidate parameter set generated by the optimizer, the closed-loop CSTR response is
simulated and its performance is evaluated using the normalized Zwe–Lee–Gaing objective function, denoted as ZLGnorm [55]. This
scalar performance index combines overshoot, steady-state error, settling time, and rise time into a unified metric and is defined as
follows:
(os ssenorm )
norm
ZLGnorm = (1 − φ) + + φ(st norm − rt norm ), (31)
100 100

where osnorm represents the normalized percent overshoot, ssenorm denotes the normalized steady-state error evaluated at t = 20 min,
and stnorm and rtnorm correspond to the normalized settling and rise times, respectively. The balance factor is selected as ϕ = e− 1 =
0.3679, ensuring a compromise between overshoot suppression and fast settling.
The gd-PID controller contains six tuneable parameters,
[ ]
Θ = Kp , Ki , Kd , N, G1 , G2 , (32)

and the constrained search domain is defined by

0.05 ≤ Kp ≤ 1, 0.1 ≤ Ki ≤ 2, 0.02 ≤ Kd ≤ 0.4,

1 ≤ N ≤ 10, 0.1 ≤ G1 ≤ 5, 0.1 ≤ G2 ≤ 5. (33)

Optimization procedure shown in Fig. 6 proceeds as follows:
(1) THRO generates a new parameter vector Θ,
(2) gd-PID controller is updated with these parameters,
(3) Nonlinear CSTR model is simulated under the specified step-change scenario,
(4) Metrics for time-domain stability (overshoot, settling time, rise time, steady-state error) are extracted,
(5) Objective function ZLGnorm is computed and returned to THRO for the next iteration.

Fig. 6. THRO-driven closed-loop optimization framework for tuning the gd-PID controller on the CSTR model.

11
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

In this study, the gd-PID parameters are not selected heuristically. All tunable parameters are obtained via offline optimization
using the THRO algorithm within an outer-loop simulation framework. At each THRO iteration, a candidate parameter vector is
injected into the nonlinear CSTR closed loop, the response is simulated under the prescribed operating scenario, and the objective
value is computed from the resulting time-domain metrics. The search is constrained within the parameter bounds defined in (33), and
the final parameter set is selected as the best solution returned by THRO after the termination criterion is met.
Here, closed-loop optimization framework constructs a nonlinear, bounded, and multidimensional search problem that allows the
gd-PID controller to be tuned using the complete dynamic behaviour of the CSTR process. Through this formulation, both stability and
transient performance are quantified, enabling a fair comparison with alternative control strategies.

6. Comparative simulation results

This section studies the tuning performance of chosen powerful methods, namely THRO [80], BKA [81], ASO [82], WOA [83], and
GA [84] under identical experimental settings of 100 iterations, a population size of 30, and 25 independent runs. The analyses include
some statistical indicators (best, worst, mean, SD), Mann–Whitney U test results, convergence trajectories, and the optimized gd-PID
parameter sets achieved by each algorithm. Time-domain performance metrics such as rise time, settling time, overshoot, and integral
absolute error are also evaluated to characterize the closed-loop behaviour of the CSTR system. The section presents these findings in a
structured manner based on the numerical and dynamic outcomes generated in the experiments.

6.1. Statistical analysis

The outcomes obtained from 25 independent executions are reported in Table 1 and further visualized in Fig. 7. Rather than
reflecting a single run, these statistics outline how consistently each algorithm performs across repeated trials. In this set of experi­
ments, THRO attains the smallest observed objective value (0.4671), indicating that it is able to reach a comparatively favourable
solution in at least one run. The results achieved by BKA and ASO are also close, although their best values remain slightly above the
minimum reached by THRO.
In terms of worst-case performance, THRO again outperforms the alternatives, maintaining the lowest maximum objective value
(0.4883). This indicates that even under less favourable conditions, THRO avoids the performance degradation observed in WOA and
GA, both of which show substantially higher worst values (0.5889 and 0.6232, respectively). The mean values reinforce this trend:
THRO attains the smallest average objective value (0.4780), confirming its overall stability across repeated trials. BKA and ASO remain
moderately stable, whereas WOA and GA exhibit larger deviations, suggesting increased sensitivity to stochastic behaviour. The
standard deviation values further underline these observations. THRO presents the smallest variability (SD = 0.0048), indicating
highly consistent convergence behaviour. BKA and ASO show slightly higher fluctuations, while WOA and GA display the largest
deviations, aligning with their broader spread in Fig. 7. The bar plot visually confirms that THRO's run-to-run dispersion is narrower
than that of the competing algorithms, which supports its superior robustness.
Overall, the combined interpretation of Table 1 and Fig. 7 demonstrates that THRO delivers the most accurate, stable, and
repeatable performance among the tested methods. The consistently lower best, worst, and mean values, together with its minimal
variance, indicate that THRO maintains strong optimization capability even under stochastic conditions.

6.2. Mann-Whitney U test

To statistically validate the performance differences observed between THRO and the competing algorithms, the non-parametric
Mann–Whitney U test is applied using a significance level of α = 0.05. The results are summarized in Table 2. The h-values are equal to
1 for all comparisons, indicating that the null hypothesis, stating that the two samples originate from identical distributions, is rejected
in every case.
The corresponding p-values reveal the magnitudes of these differences. Comparisons between THRO and WOA, as well as THRO
and GA, yield extremely small p-values on the order of 10− 9, indicating strong evidence that THRO exhibits statistically superior
performance relative to these methods. The THRO–BKA comparison also shows a highly significant difference (p = 7.6757 × 10− 8 ),
confirming that the superiority of THRO is not due to random variation. Even in the case of ASO, which is relatively closer to THRO in
raw performance, the p-value of 0.0189 remains below the significance threshold. Overall, the Mann–Whitney results demonstrate that
the performance improvements achieved by THRO are statistically significant against all benchmarked algorithms, reinforcing the
conclusions drawn from the descriptive statistics and run-to-run analysis.

Table 1
Statistical performance indicators (best, worst, mean, SD) of THRO, BKA, ASO, WOA, and GA across 25 optimization runs.
Statistical measure THRO BKA ASO WOA GA

Best 0.4671 0.4784 0.4745 0.5303 0.5536

Worst 0.4883 0.5068 0.5017 0.5889 0.6232
Mean 0.4780 0.4899 0.4822 0.5485 0.5748
SD 0.0048 0.0066 0.0067 0.0124 0.0157

12
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

Fig. 7. Statistical comparison of optimization results across 25 independent runs.

Table 2
Mann–Whitney U test results comparing THRO with BKA, ASO, WOA, and GA under a 5% significance level.
Metric THRO vs. BKA THRO vs. ASO THRO vs. WOA THRO vs. GA

p-value 7.6757E− 08 0.0189 1.4144E− 09 1.4144E− 09

h 1 1 1 1

6.3. Evolution of objective function and optimized gd-PID controller parameters

Graphic visualization of the convergence behaviour of the algorithms is illustrated in Fig. 8, presenting a detailed view of how
rapidly and successfully each method reduces the objective function value over the 100 iterations. THRO exhibits the steepest initial
descent, reaching a value close to 0.60 within the first 10 iterations and continuing to improve steadily until convergence. This early-
stage reduction demonstrates THRO's strong global search capability, while its subsequent fine-tuning phase indicates efficient local
refinement. BKA and ASO also achieve relatively fast convergence, although both remain consistently above THRO across the entire
optimization horizon. In contrast, WOA and GA converge more slowly, displaying extended plateaus and larger fluctuations, which
reflects weaker local exploitation and a tendency to become trapped in suboptimal regions.
Optimized gd-PID parameters obtained by each algorithm are listed in Table 3. These results reveal clear structural differences in
how the methods explore and tune the parameter space. THRO produces the highest proportional gain (Kp = 0.9312), suggesting that
it favors stronger corrective action nearby the reference. It also identifies a relatively large integral gain (Ki = 1.8416), supporting
improved steady-state accuracy. Compared to the other algorithms, THRO selects a moderate derivative gain (Kd = 0.2150),

Fig. 8. Convergence profiles of THRO, BKA, ASO, WOA, and GA over 100 iterations.

13
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

indicating an effective balance between noise sensitivity and transient shaping.

The nonlinear scaling parameters (G1 and G2 ) further highlight the distinction between THRO and the competing methods. THRO
yields the largest G2 value (4.7349), implying a more pronounced utilization of the Gudermannian-based nonlinear transformation to
modulate the error signal. Meanwhile, its chosen filter bandwidth (N = 8.3705) places the derivative action in a stable and well-
damped regime. Other algorithms present more conservative choices of G1 and G2 , with ASO and WOA selecting comparatively
lower values, which may limit the nonlinear sensitivity benefits provided by the gd-PID structure.
Overall, the convergence trajectories and the optimized parameter set consistently indicate that THRO identifies more effective gd-
PID configurations compared to BKA, ASO, WOA, and GA. The parameter combinations generated by THRO align with the observed
performance improvements in the earlier sections, confirming its superior ability to exploit both the linear and nonlinear components
of the controller architecture.

6.4. Time response and step-response characteristics

The closed-loop reactor temperature responses obtained using the optimized gd-PID controllers are presented in Fig. 9. The
reference profile consists of a 10 K step change applied at t = 1 min:
{
304.16755 K if 0 ≤ t < 1 min
T= (34)
314.16755 K if t ≥ 1 min

and all algorithms are evaluated for their capability to track this set-point under nonlinear CSTR dynamics. As shown in Fig. 9, THRO-
based gd-PID exhibits the fastest and most stable convergence to the new operating temperature. Its response rapidly approaches the
reference level with minimal overshoot and no observable oscillations, indicating that the parameter set identified by THRO provides
an effective balance between transient aggressiveness and damping.
BKA- and GA-based controllers follow with relatively smooth trajectories but display slightly longer rise times. ASO and WOA show
noticeably slower responses and more pronounced deviation during the early transient period, reflecting weaker tuning of the
nonlinear scaling parameters and derivative filtering. These behaviours are consistent with the convergence profiles presented earlier,
where THRO achieved the most effective optimization of control parameters.
To complement the qualitative time-domain observations, additional quantitative performance indices are employed to provide an
objective assessment of tracking accuracy and control effort. In addition to rise time, settling time, and overshoot, which are directly
reported in Fig. 10, several standard integral performance measures are considered over the evaluation horizon tf = 20 min:

∫tf
ISE = e2 (t)dt (35)
0

∫tf
ITAE = t|e(t) |dt (36)
0

∫tf
ITSE = te2 (t)dt (37)
0

∫tf
Eng = u2 (t)dt (38)
0

where e(t) denotes the reactor temperature tracking error and u(t) represents the manipulated jacket temperature. The ISE metric
quantifies cumulative tracking error, the ITAE and ITSE metrics impose time-weighted error penalties, and the energy index (Eng)
reflects the control effort required by each controller. The resulting numerical values for all optimization algorithms are summarized in
Table 4.

Table 3
Optimal gd-PID parameters obtained by THRO, BKA, ASO, WOA ad GA.
Parameter THRO BKA ASO WOA GA

Kp 0.9312 0.7045 0.5485 0.6440 0.8371

Ki 1.8416 1.5318 1.8518 1.9408 1.5502
Kd 0.2150 0.2342 0.1987 0.3922 0.3478
N 8.3705 7.3830 7.7928 9.2307 6.7604
G1 4.3449 4.5452 2.9765 3.0291 3.5972
G2 4.7349 4.1841 4.2721 4.2656 4.3258

14
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

Fig. 9. Reactor-temperature responses produced by gd-PID controllers tuned with different algorithms.

Fig. 10. Comparative rise time, settling time, overshoot, and IAE metrics for optimized controllers.

Table 4
Integral performance indices and control effort for gd-PID controllers tuned with different optimization algorithms.
Performance indices THRO BKA ASO WOA GA

ISE 15.1421 24.1673 26.5788 15.0485 15.7078

ITAE 3.6023 6.3252 6.6462 5.0171 4.6189
ITSE 17.4601 30.3243 33.3759 18.1380 18.6257
Eng 1.6946E+06 1.6936E+06 1.6935E+06 1.6944E+06 1.6944E+06

15
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

A more detailed comparison of dynamic metrics is given in Fig. 10, which reports the rise time, settling time, overshoot, and in­
tegral absolute error (IAE). Formula of the IAE metric is defined as;

∫tf
IAE = |e(t) |dt, tf = 20 min (39)
0

Since the error is expressed in Kelvin and time is expressed in minutes in this study, the IAE metric has units of Kelvin–minute
(K⋅min). When the four performance indicators are considered together, THRO shows a clear advantage. It records the shortest rise
time (0.5861 min) and reaches steady state more rapidly, with a settling time of 1.8521 min, while also limiting overshoot to 0.2186%.
In addition, the IAE obtained with THRO (2.8580) remains lower than those reported for the other methods, reflecting a more effective
reduction of accumulated error over the 20-min evaluation window. Taken as a whole, these observations indicate that THRO
maintains favourable transient characteristics alongside improved tracking accuracy over longer time horizons.
By comparison, ASO and WOA generate noticeably higher overshoot levels, reaching 1.2121% and 1.3039%, respectively. BKA and
GA exhibit less pronounced overshoot, yet their responses converge more slowly. For all competing algorithms, the settling times
exceed those achieved by THRO, and the corresponding IAE values remain higher, suggesting weaker disturbance handling and greater
cumulative error.
With this, the time-domain results demonstrate that THRO achieves the most desirable combination of rapid convergence, low
overshoot, and minimal cumulative error, validating its advantage in tuning the gd-PID controller for the nonlinear CSTR system.

6.5. Comparison with well-known control methods

The performance of the proposed gd-PID controller is evaluated against four widely used structures: PI, PIDf, 2-DOF PID and FOPID.
All controllers are tuned using the THRO algorithm under identical constraints and tested on the same CSTR temperature-tracking
scenario. Their mathematical forms and optimized parameter values are presented below, followed by a detailed comparative anal­
ysis based on the step-response curves in Fig. 11 and the numerical performance indices in Fig. 12.
The PI controller is defined in Laplace (s) domain as [85]:
( )
Ki
U(s) = E(s) Kp + , (40)
s

where Kp and Ki denote proportional and integral gains. THRO tuning yielded;
Kp = 0.2516, Ki = 1.7850. (41)

The filtered PID (PIDf) structure in Laplace (s) domain is expressed as [86,87]:
( )
Ki Ns
U(s) = E(s) Kp + + Kd , (42)
s s+N

where Kd is the derivative gain and N is the filter coefficient. The optimized parameters are
Kp = 0.3680, Ki = 1.8177, Kd = 0.0389, N = 4.3212. (43)

Fig. 11. Temperature-tracking response of PI, PIDf, 2-DOF PID, FOPID, and gd-PID structures.

16
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

Fig. 12. Performance indices summarizing step-response behaviour across the five controller designs, including IAE.

The two-degree-of-freedom PID (2-DOF PID) controller in Laplace (s) domain is modelled as [88,89];
Ki Ns
U(s) = Kp [bR(s) − Y(s) ] + E(s) + Kd [cR(s) − Y(s) ], (44)
s s+N

where b and c represent the proportional and derivative set-point weightings. The THRO-based parameters are;
Kp = 0.5526, Ki = 1.5791, Kd = 0.1198, N = 7.8912, b = 1.3746, c = 0.7308. (45)

And finally, the fractional-order PID (FOPID) controller in Laplace (s) domain is defined by [88,90,91];
( )
U(s) = E(s) Kp + Ki s− λ + Kd sμ , (46)

where λ and μ denote fractional integral and derivative orders. The optimized parameter set is
Kp = 0.9961, Ki = 1.9140, Kd = 0.0772, λ = 1.0392, μ = 1.0786. (47)

Inspection of Fig. 11 suggests clear differences in the transient responses. The gd-PID response rises more rapidly at the initial stage
and approaches the reference earlier than the other controllers. This behaviour can be associated with the selected parameter set,
which strengthens the corrective action during the early phase without leading to instability. PI and PIDf responses evolve more
slowly, resulting in a prolonged temperature increase over roughly the first 2 min. Although the 2-DOF PID and FOPID yield smoother
trajectories, their convergence remains slower than that observed with gd-PID. The relatively steeper response of these alternatives
during the first minute indicates a weaker initial drive, which can be limiting in CSTR applications where prompt temperature
regulation is required.
The numerical results reported in Fig. 12 support the trends observed in the transient responses. The rise time achieved by the gd-
PID controller (0.5861 min) is approximately half of that obtained with PI control (1.1621 min) and remains lower than those recorded
for PIDf (1.2814 min), 2-DOF PID (1.4422 min), and FOPID (1.3546 min). This reduction in rise time suggests that the gd-PID can
approach updated operating setpoints more rapidly, thereby limiting the duration of intermediate thermal states. Such behaviour is
relevant in temperature-sensitive or exothermic reaction regimes, where extended transients may be undesirable.
Differences among the controllers become more apparent when the settling time is examined. The gd-PID reaches steady operation
in 1.8521 min, whereas PI and PIDf require 3.9212 min and 3.8653 min, respectively. Although the 2-DOF PID (3.0758 min) and
FOPID (3.0505 min) improve upon classical structures, their stabilization times remain longer than that of the gd-PID. This behaviour
indicates that the proposed controller approaches the final value with fewer sustained oscillations or recovery phases. From an
operational perspective, shorter settling intervals can reduce cumulative energy usage and limit exposure to off-target temperature
levels.

17
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

Overshoot behaviour reveals another significant distinction. PI and PIDf overshoot by 3.7056% and 2.4677%, respectively,
reflecting their limited damping ability. The 2-DOF PID and FOPID structures, designed to improve overshoot control, reduce the peak
to 0.8339% and 0.9665%. However, the gd-PID achieves an exceptionally small overshoot of 0.2186%. This value is nearly negligible
for CSTR applications and demonstrates that the gd-PID ensures both fast and safe temperature transitions. Excessive overshoot in
reactor control can cause undesirable thermal acceleration, catalyst degradation, or reaction-rate shifts; thus, the gd-PID's ability to
minimize it represents a crucial operational advantage.
The IAE metric provides a cumulative error assessment, capturing the accuracy of the entire transient and steady-state behaviour.
The gd-PID records an IAE of 2.8580, which is dramatically lower than PI (7.3132), PIDf (6.9575), 2-DOF PID (6.2530) and FOPID
(6.3752). This indicates that the proposed method produces far smaller sustained deviations from the reference signal. Such error
reduction translates directly to improved control efficiency and more precise thermal regulation throughout the process duration.
Taken together, the complete set of results clearly demonstrates that the gd-PID controller surpasses all conventional and advanced
counterparts across every key performance indicator, including rise time, settling time, overshoot and IAE. While 2-DOF PID and
FOPID generally outperform classical PI-type controllers in many applications, the THRO-optimized gd-PID structure provides a su­
perior balance of speed, damping and accuracy. This performance can be attributed to the proposed error-shaping mechanism, which
effectively regulates the balance between responsiveness and stability. In particular, the attenuation of large error magnitudes reduces
aggressive control actions, while preserving sensitivity near the operating point, thereby explaining the improved transient behaviour
observed in the gd-PID responses. Consequently, the gd-PID emerges as the most effective control strategy for the investigated CSTR
temperature-tracking problem.

7. Robustness of THRO-based gd-PID controller

From a nonlinear systems perspective, the proposed gd-PID controller does not attempt to cancel or linearize the inherent non­
linearities of the CSTR. Instead, the nonlinear reaction–temperature coupling governed by Arrhenius kinetics remains fully active in
the closed-loop dynamics, while the control action is reshaped through an error-dependent feedback mechanism. This results in a state-
dependent closed-loop behaviour rather than a disturbance-rejection-only structure. In particular, the bounded and smoothly varying
slope of the Gudermannian mapping introduces a nonlinear modulation of the effective loop gain. As the system deviates from the
operating point, the equivalent feedback gain decreases progressively, enhancing damping during large thermal excursions, while
preserving high sensitivity near equilibrium conditions. This mechanism provides a qualitative nonlinear interpretation of how the gd-
PID controller alters the closed-loop behaviour compared to conventional PID structures without overriding the intrinsic plant
dynamics.
It is noted that the present robustness analysis is conducted around a nominally stable operating point of the CSTR, which is
representative of typical industrial operation. The objective of the proposed gd-PID framework is not to suppress or linearize the
inherent nonlinearities of the reactor, but to reshape the closed-loop error dynamics through bounded and state-dependent gain
modulation. While a formal bifurcation or unstable steady-state analysis is beyond the scope of this study, the analytic and bounded
nature of the Gudermannian-based error shaping inherently limits excessive control action as the operating regime deviates from
nominal conditions. This structural property suggests potential suitability for operation near critical regimes, which is identified as an
important direction for future investigation. In addition, the bounded and smooth nature of the proposed error transformation reduces
the sensitivity of the closed-loop response to parameter variations and tuning inaccuracies, thereby limiting the risk of instability under
uncertain operating conditions.
The robustness of the proposed THRO-based gd-PID controller is further investigated by examining the behaviour of the system
states and manipulated variable under a step change in the temperature reference. In addition to evaluating the ability of the controller
to shape the reactor temperature dynamics, it is crucial to assess whether the internal variables of the CSTR, particularly the jacket
temperature Tj and the concentration of species A, denoted CA , remain well behaved during transient phases and smoothly converge to
new operating conditions. These internal responses provide a deeper understanding of control robustness, actuator safety, thermo­
dynamic coherence, and dynamic consistency within the closed loop.
In the following subsections, the variations in Tj (t) and CA (t) are first analysed, after which the controller's reference-tracking
capability, disturbance-rejection performance, and noise-attenuation characteristics will be examined.

7.1. Controller output: Tj and CA responses

A 10-K temperature increase is applied at t = 1 min, shifting the reactor temperature setpoint from T = 304.16755 K to T =
314.16755 K. Fig. 13 illustrates the corresponding evolution of the jacket temperature Tj (t), which also represents the control signal of
the gd-PID controller. The saturation limits for the actuator are imposed as 200 ≤ Tj ≤ 400 K; throughout the entire operation, the
control signal remained strictly within these bounds without exhibiting clipping or artificial flattening. This indicates that the
controller achieves the required thermal compensation through dynamically consistent and physically feasible inputs.
From a practical standpoint, this bounded and smooth actuator behaviour implies that the proposed controller avoids excessive
control effort and actuator saturation while maintaining consistent thermal regulation. Such characteristics are essential for real-world
applications, where actuator limitations, energy efficiency, and operational safety must be considered.
Immediately after the reference step, the jacket temperature rises sharply to provide the required heat input to the reactor. The
initial peak around Tj ≈ 300 K reflects a strong but well-regulated heating action. A short sequence of damped oscillatory behaviour

18
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

follows, caused by the rapid change in thermal load and the nonlinear energy balance of the CSTR system. These oscillations remain
bounded and decay within approximately 1 min, demonstrating that the gd-PID structure possesses adequate damping and does not
induce actuator instability. After the transient phase, Tj (t) smoothly converges to a new steady-state value of:

Tj,∞ = 291.5790 K. (48)

This value is thermodynamically consistent with the increased reactor temperature, reflecting the shift in the heat-removal
requirement caused by the new operating point.
The closed-loop behaviour of the reactant concentration CA (t) is presented in Fig. 14. Prior to the setpoint change, the concen­
tration remains nearly constant. Once the reactor temperature increases, the reaction rate accelerates according to the Arrhenius
relationship, leading to a natural decline in the concentration of species A. This temperature-induced kinetic shift is visible as a smooth
downward trajectory starting shortly after the 1-min mark. Unlike the jacket temperature response, the concentration profile exhibits
no oscillation or abrupt curvature, indicating that the internal chemical dynamics remain fully stable under the gd-PID-regulated heat
input. The concentration progressively settles to:
/
CA,∞ = 0.9454 mol L, (49)

representing the new equilibrium operating condition of the reactor under the elevated temperature.
The combined evaluation of Fig. 13 and Fig. 14 shows that the THRO-based gd-PID controller ensures a well-conditioned transient
in both the manipulated variable and process states. The jacket temperature responds aggressively enough to track the new setpoint
while remaining within physical limits, and the concentration profile evolves smoothly without any undesirable dynamic excursions.
These results affirm the robustness of the proposed controller when subjected to significant temperature transitions, providing a solid
basis for evaluating its performance under more demanding scenarios such as reference tracking, external disturbances, and mea­
surement noise, which will be examined in the subsequent subsections.

7.2. Tracking of reference signal

To further assess the dynamic tracking capability of the THRO-based gd-PID controller, the reactor temperature is subjected to a
piecewise-constant reference profile consisting of multiple upward and downward transitions of varying magnitudes. The reference
trajectory is defined as
⎧
⎪
⎪ 304.16755 K if 0 ≤ t < 1 min
⎪
⎪
⎪
⎪ 319.16755 K if 1 ≤ t < 13 min
⎨
294.16755 K if 13 ≤ t < 26 min
T= (50)
⎪
⎪ 309.16755 K if 26 ≤ t < 37 min
⎪
⎪
⎪
⎪ 314.16755 K if 37 ≤ t < 48 min
⎩
304.16755 K if 48 ≤ t < 60 min

The multi-step reference trajectory in (50) is selected to reflect realistic operating conditions in which the reactor temperature
setpoint is updated in discrete stages rather than as a single large jump. This design enables the controller to be evaluated under both
upward and downward transitions with different amplitudes, allowing tracking capability and transient consistency to be assessed
across multiple operating regions. In addition, the dwell times between successive steps are chosen to ensure that the response has

Fig. 13. Time response of jacket temperature (Tj ) for THRO-based gd-PID controlled system.

19
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

Fig. 14. Time response of concentration of A in reactor (CA ) for THRO-based gd-PID controlled system.

sufficient time to approach the new steady operating condition, preventing the performance indices from being biased by incomplete
settling. The same reference profile is applied to all compared controllers to preserve fairness in the benchmarking framework.
The closed-loop reactor temperature response is shown in Fig. 15. The gd-PID controller demonstrates excellent tracking accuracy
throughout the entire 60-min interval, responding rapidly to each reference transition while preserving stability and smoothness.
During the first step at t = 1 min, the reactor temperature rises sharply and settles near the new setpoint with minimal overshoot,
which is consistent with the fast transient performance observed in earlier analyses. The close overlap between the reference trajectory
and the controlled response confirms that the controller provides strong proportional dominance with effective integral compensation,
enabling the system to quickly eliminate steady-state error.
The second transition, occurring at t = 13 min and representing a significant downward step to 294.16755 K, poses a more
challenging thermal demand because the reactor must rapidly dissipate heat while maintaining reaction stability. The gd-PID
controller handles this transition smoothly: although the controlled temperature temporarily drops slightly below the setpoint due
to the nonlinear energy balance of the CSTR, it rapidly recovers without oscillatory behaviour. This behaviour suggests that the
controller maintains sufficient damping even during aggressive cooling phases.
The subsequent upward transition at t = 26 min again highlights the agility of the closed loop. The reactor temperature rises
quickly to 309.16755 K, and the tracking error diminishes in a short time, demonstrating that the controller effectively adapts to both
heating and cooling directions without bias. The same trend is observed in the rise to 314.16755 K at t = 37 min, where the gd-PID
provides a near-perfect match to the reference.
Finally, the transition back to the nominal temperature of 304.16755 K at t = 48 min shows that the controller remains stable even
after multiple successive setpoint changes. The response settles into the final value without oscillation or drift, indicating that neither
actuator fatigue nor integral windup has accumulated during the sequence of transitions.
Overall, Fig. 15 demonstrates that the THRO-based gd-PID controller exhibits a high level of robustness to varying reference

Fig. 15. Multi-step reference-tracking performance of the THRO-based gd-PID controller.

20
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

conditions, providing fast rise times, smooth convergence, negligible steady-state error, and consistent tracking performance across a
wide range of temperature shifts. The results confirm that the controller maintains its dynamic quality not only under single-step
changes but also under complex multi-step reference trajectories that better reflect realistic operating conditions in industrial CSTR
processes.

7.3. Disturbance rejection performance

The disturbance-rejection performance of the THRO-based gd-PID controller is evaluated by introducing a sequence of feed-
temperature variations, representing one of the most influential disturbances in non-isothermal CSTR operations. The disturbance
signal applied to the system is defined as:
⎧
⎪
⎪ 350 K if 0 ≤ t < 5 min
⎨
380 K if 5 ≤ t < 15 min
Tf = (51)
⎪
⎪ 330 K if 15 ≤ t < 22 min
⎩
350 K if 22 ≤ t < 30 min

Fig. 16 illustrates this disturbance profile. The magnitude and direction of the variations are selected to represent realistic changes
in inlet stream conditions arising from upstream preprocessing fluctuations, ambient thermal drift, or mixing-stage variability.
The resulting reactor temperature response is shown in Fig. 17. After the initial tracking of the reference value around 324 K, the
first disturbance occurs at t = 5 min when the feed temperature abruptly increases from 350 K to 380 K. This disturbance instanta­
neously increases the sensible enthalpy entering the reactor, creating a positive thermal imbalance. The closed-loop response exhibits a
small but noticeable temperature elevation above the reference, peaking slightly above the setpoint. However, the deviation is short-
lived: the gd-PID controller immediately compensates by adjusting the jacket temperature downward, leading the system back to the
reference with a smooth and well-damped convergence. The rapid correction in this phase demonstrates the controller's ability to
eliminate the influence of strong and sudden thermal disturbances within a narrow time window.
At t = 15 min, the feed temperature drops drastically to 330 K, corresponding to a negative disturbance of the same magnitude.
This shift causes the energy entering the reactor to momentarily fall below the level required to maintain the reference temperature. As
expected, the reactor temperature briefly dips significantly below the setpoint. This downward deviation is visibly larger than the
initial upward deviation, which can be attributed to the stronger net cooling effect and nonlinear heat-transfer behaviour of the CSTR
around lower feed temperatures. Nonetheless, the gd-PID controller promptly reacts by raising the jacket temperature, counteracting
the cooling disturbance. The system returns to the reference trajectory without oscillatory behaviour, confirming that the controller
maintains stability even under severe negative disturbances.
A third disturbance occurs at t = 22 min, where the feed temperature is restored from 330 K to 350 K. The reactor temperature
exhibits a small overshoot during this upward correction, again due to the nonlinear enthalpy response inherent in the reactor model.
However, this overshoot is significantly smaller than would be expected under a conventional PID structure, and the gd-PID ensures
rapid re-stabilization after the deviation.
Across all disturbance intervals, the controller maintains a strong ability to reject thermal perturbations while preserving the
desired temperature trajectory. The deviations remain bounded, short in duration, and free from sustained oscillation or drift. The
reactor returns to the reference value within a few minutes after each disturbance, demonstrating that the gd-PID gains tuned via
THRO provide a robust compensation mechanism capable of handling both positive and negative enthalpy shocks. These results
confirm the controller's suitability for practical CSTR environments where feed-temperature fluctuations are common and often
unavoidable.

Fig. 16. Applied feed-temperature disturbance sequence used in the robustness tests.

21
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

Fig. 17. Reactor-temperature response illustrating disturbance-rejection capability.

7.4. Noise suppression performance

In practical CSTR operations, measurement noise is unavoidable due to sensor resolution limits, electronic interference, thermal
fluctuations, and stochastic disturbances in the measurement environment. To examine the noise-suppression performance of the
THRO-based gd-PID controller, band-limited white noise is injected into the reactor temperature measurement during closed-loop
operation. The noise characteristics are defined as a noise power of 0.0004 W and a sample time of 0.05 min. The applied noise
sequence is illustrated in Fig. 18, showing rapid and irregular fluctuations with amplitudes ranging approximately from − 0.25 K to
+0.25 K.
The resulting controlled reactor-temperature response is presented in Fig. 19. Despite the relatively strong noise content, the gd-
PID controller successfully preserves the overall shape and stability of the tracking response. Following the reference transition, the
reactor temperature rises smoothly toward the 324 K setpoint, and the presence of noise does not significantly affect the rise-time or
settling-time characteristics. In steady state, the temperature fluctuations remain confined within a narrow band around the reference
value, with an approximate peak-to-peak variation of 0.532 K, corresponding to an operating interval of roughly [323.8 K, 324.3 K].
This behaviour indicates that high-frequency measurement noise is effectively attenuated by the closed-loop system and does not
induce amplification or oscillatory drift. The transient response therefore remains comparable to the noise-free case, confirming the
robustness of the gd-PID control action under noisy operating conditions.
After the system reaches steady-state conditions, the noise-induced fluctuations in the measured temperature manifest as small-
amplitude ripples around the desired operating point. Importantly, these ripples remain extremely limited in magnitude: the
closed-loop temperature oscillates only within a narrow band around the reference value, indicating that the control law does not
amplify high-frequency sensor noise. This behaviour demonstrates that the controller's derivative-free structure, in conjunction with
the THRO-tuned gains, inherently provides a filtering effect that mitigates measurement disturbances. In contrast to classical PID
controllers whose derivative term often increases vulnerability to noise amplification, gd-PID structure avoids introducing sharp
control actions that could drive instability or excessive actuator motion.
The jacket-temperature response follows a smooth trajectory over the simulated interval, indicating that the manipulated input
does not exhibit noticeable jitter or chattering. Under these conditions, actuator activity remains moderate, which can help limit
unnecessary wear and excessive energy use. From a practical standpoint, avoiding rapid oscillations in heating or cooling commands is
advantageous in thermal processes, as repeated high-frequency variations may otherwise introduce additional thermal stress or
shorten actuator service life over extended operation.
Fig. 18 and Fig. 19 illustrate the response of the THRO-tuned gd-PID controller under measurement noise. Across the tested cases,
the controller continues to operate stably in the presence of persistent high-frequency disturbances, while overshoot levels and
tracking accuracy remain largely unaffected. This behaviour suggests that the proposed approach can tolerate noisy measurements in
addition to reference changes and external disturbances, a condition that commonly arises in practical reactor operation.

8. Conclusion and potential future works

This study considers a modified PID control formulation in which the Gudermannian function is introduced into the error-
processing stage, with controller parameters adjusted through Tianji's Horse Racing Optimizer (THRO). The gd-PID structure is
examined on a nonlinear continuous stirred-tank reactor (CSTR), selected mainly due to its thermal nonlinearity and sensitivity to
operating conditions. Instead of redefining the entire controller architecture, the approach focuses on altering how the error signal is
transformed, which affects the resulting control action in a less direct but noticeable manner. The intent is not to replace classical PID
behaviour, but to adjust its response characteristics under conditions where standard formulations may become less predictable.
The simulation results indicate that incorporating the Gudermannian function leads to a more regulated control response,

22
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

Fig. 18. Band-limited white noise injected into the reactor-temperature measurement.

Fig. 19. Controlled reactor-temperature response under measurement-noise conditions.

particularly during transient phases. In the tested scenarios, the gd-PID controller shows faster rise and settling characteristics
compared to PI, PIDf, 2-DOF PID, and FOPID structures, while overshoot and accumulated error levels remain comparatively low.
These trends suggest that the modified error transformation helps balance early corrective action with stability, without introducing
discontinuities or abrupt gain changes.
Internal system responses further support these observations. Jacket temperature and reactant concentration trajectories remain
physically consistent during operating-point transitions, and no unstable or unrealistic behaviour is observed in the closed-loop
simulations. While the improvements are most visible during large reference changes, steady-state behaviour remains comparable
to that of classical PID-based designs.
Additional tests under varying operating conditions provide insight into the controller's behaviour beyond nominal cases. Refer­
ence tracking remains stable under multi-level setpoint changes, and feed-temperature disturbances are attenuated within a short time
window. The response to measurement noise does not show excessive amplification of high-frequency components, which indicates
that the nonlinear shaping introduced by the Gudermannian mapping does not adversely affect noise sensitivity.
While the proposed gd-PID framework is formulated with practical implementation considerations in mind, the present study
focuses on a generalized application framework rather than deployment on a specific reactor installation. The controller design
intentionally preserves the classical PID structure and relies on standard process measurements and a physically realizable manipu­
lated variable, ensuring compatibility with conventional industrial control hardware. From a computational perspective, the online
execution burden remains equivalent to that of a standard PID controller, since the Gudermannian error shaping is evaluated in closed
form and does not introduce iterative or recursive operations. Parameter optimization is performed offline, thereby avoiding addi­
tional real-time computational load.
Measurement noise and operating variability are addressed at the evaluation level through disturbance scenarios and noise-injected
simulations, which reflect typical sensing conditions encountered in industrial environments. At the same time, hardware-specific
effects such as actuator saturation limits, sensor dynamics, communication delays, and platform-dependent sampling constraints

23
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

are not modelled explicitly, as these factors vary across reactor configurations and control infrastructures. Rather than tailoring the
controller to a single hardware setup, the adopted benchmarking-oriented strategy prioritizes structural comparison, reproducibility,
and controller behaviour under representative operating conditions. As a result, the reported findings are intended to capture general
performance trends and implementation-relevant characteristics, rather than device-specific outcomes tied to a particular experi­
mental platform.
The nonlinearity introduced by the Gudermannian error-shaping function does not impose a significant computational overhead in
practical implementation. The transformation is evaluated in closed form using a fixed number of elementary operations and does not
involve iterative computation, adaptive updating, or additional state variables. Consequently, the online execution complexity of the
gd-PID controller remains comparable to that of a conventional PID controller. This property supports the suitability of the proposed
approach for real-time industrial control environments where deterministic execution time is required.
Building on the results reported in this study, several directions naturally follow from the proposed framework. While THRO is
employed here as the tuning mechanism, the gd-PID structure may also be examined in conjunction with alternative or hybrid
optimization strategies to further explore its interaction with different search dynamics. The integration of anti-windup schemes or
constraint-aware formulations represents another potential extension, particularly in applications where actuator limits become
active.
Beyond the CSTR case considered in this work, it is emphasized that the proposed gd-PID framework is not inherently tied to a
specific process model, since the Gudermannian-based error shaping operates solely on the control error rather than on plant-
dependent dynamics. Accordingly, applying the gd-PID structure to other nonlinear thermal or chemical processes, such as
exothermic reactors or heat-exchanger networks, could help delineate the broader applicability of the proposed error-shaping concept.
In addition, hardware-in-the-loop evaluation or pilot-scale implementation studies may complement the present simulation-based
analysis by providing further insight into platform-dependent implementation aspects. Future work may also consider the exten­
sion of the proposed gd-PID framework to fault-tolerant control scenarios, including the presence of sensor and actuator faults, to
further evaluate its robustness under abnormal operating conditions.

CRediT authorship contribution statement

Serdar Ekinci: Writing – original draft, Validation, Software, Methodology. Gokhan Yuksek: Writing – original draft, Validation,
Software. Ridvan Firat Cinar: Writing – review & editing, Validation, Software. Cebrail Turkeri: Writing – review & editing,
Validation, Software, Formal analysis. Dacheng Li: Resources, Formal analysis.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

Data availability

Data will be made available on request.

References

[1] Cherkasov N, Adams SJ, Bainbridge EGA, Thornton JAM. Continuous stirred tank reactors in fine chemical synthesis for efficient mixing, solids-handling, and
rapid scale-up. React Chem Eng Jan. 2023;8(2):266–77. [Link]
[2] Uppal A, Ray WH, Poore AB. On the dynamic behavior of continuous stirred tank reactors. Chem Eng Sci Apr. 1974;29(4):967–85. [Link]
0009-2509(74)80089-8.
[3] Thornhill NF, Patwardhan SC, Shah SL. A continuous stirred tank heater simulation model with applications. J Process Control Mar. 2008;18(3–4):347–60.
[Link]
[4] Russo LP, Bequette BW. Impact of process design on the multiplicity behavior of a jacketed exothermic CSTR. AICHE J Jan. 1995;41(1):135–47. [Link]
10.1002/aic.690410113.
[5] Kumar VR, Jayaraman VK, Kulkarni BD, Doraiswamy LK. Dynamic behaviour of coupled CSTRs operating under different conditions. Chem Eng Sci Jan. 1983;
38(5):673–86. [Link]
[6] Bequette BW. Process dynamics: modeling, analysis, and simulation. In: in Prentice-Hall international series in the physical and chemical engineering sciences.
International ed. Prentice Hall International; 1998 Accessed: Dec. 11, 2025. [Online]. Available: [Link]
[7] Henson M, Seborg D. Nonlinear process control. Accessed: Dec. 11, 2025. [Online]. Available: [Link] 1997.
[8] Dormido S, Visioli A. PID Control. In: Encyclopedia of systems and control. Cham: Springer International Publishing; 2021. p. 1724–33. [Link]
10.1007/978-3-030-44184-5_245.
[9] Johnson MA, et al. PID Control. London: Springer-Verlag; 2005. [Link]
[10] Dubey V, Goud H, Sharma PC. Role of PID control techniques in process control system: A review. In: Lecture notes in networks and systems. vol. 238.
Singapore: Springer; 2022. p. 659–70. [Link]
[11] Vlahakis E, Halikias G. Temperature and concentration control of exothermic chemical processes in continuous stirred tank reactors. Trans Inst Meas Control
Nov. 2019;41(15):4274–84. [Link]
[12] Singh A, Sharma V. Concentration control of CSTR through fractional order PID controller by using soft techniques. In: In 2013 fourth international conference
on computing, communications and networking technologies (ICCCNT). IEEE; Jul. 2013. p. 1–6. [Link]
[13] Hassan M, Deifalla H, Abdalla K. Performance evaluation of PID tuning techniques for CSTR temperature control: Classical and evolutionary approaches.
J Karary Univ Eng Sci Sep. 2025. Accessed: Dec. 11, 2025. [Online]. Available: [Link]
[14] Padula F, Visioli A. Tuning rules for optimal PID and fractional-order PID controllers. J Process Control Jan. 2011;21(1):69–81. [Link]
jprocont.2010.10.006.

24
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

[15] Chen Y, Petras I, Xue D. Fractional order control - a tutorial. In: In 2009 American control conference. IEEE; Jun. 2009. p. 1397–411. [Link]
ACC.2009.5160719.
[16] Ekinci S, Izci D, Jabari M, Ma’arif A. Enhanced temperature control of continuous stirred tank reactors using QIO-based 2-DoF PID controller. J Robot Control
(JRC) May 2025;6(3):1340–6. [Link]
[17] Das D, Chakraborty S, Naskar AK. Controller design on a new 2DOF PID structure for different processes having integrating nature for both the step and ramp
type of signals. Int J Syst Sci May 2023;54(7):1423–50. [Link]
[18] Ekinci S, Izci D, Gider V, Bajaj M, Blazek V, Prokop L. Quadratic interpolation optimization-based 2DoF-PID controller design for highly nonlinear continuous
stirred-tank heater process. Sci Rep May 2025;15(1):16324. [Link]
[19] Isaksson AJ, Graebe SF. Derivative filter is an integral part of PID design. IEE Proc Control Theory Appl Jan. 2002;149(1):41–5. [Link]
20020111.
[20] Singha P, Meena R, Chakraborty S, Raja GL. Robust pidf-pid cascade control scheme for delay-dominant stable and integrating chemical processes. J Taiwan Inst
Chem Eng Aug. 2025;173:106165. [Link]
[21] Wu JC, Liu TS. A sliding-mode approach to fuzzy control design. IEEE Trans Control Syst Technol Mar. 1996;4(2):141–51. [Link]
[22] Tang KS, Man Kim Fung, Chen Guanrong, Kwong S. An optimal fuzzy PID controller. IEEE Trans Ind Electron Aug. 2001;48(4):757–65. [Link]
10.1109/41.937407.
[23] Carvajal J. Fuzzy PID controller: design, performance evaluation, and stability analysis. Inform Sci (N Y) Apr. 2000;123(3–4):249–70. [Link]
S0020-0255(99)00127-9.
[24] Camacho O, Vega S, Herrera M, Di Teodoro A, Gude JJ. A fractional order PID-based sliding mode controller approach for chemical processes. Res Control
Optim Sep. 2025;20. [Link]
[25] Li M, Wang F, Gao F. PID-based sliding mode controller for nonlinear processes. Ind Eng Chem Res Jun. 2001;40(12):2660–7. [Link]
IE990715E.
[26] Huba M, Bistak P, Vrancic D. Series PIDA controller design for IPDT processes. Appl Sci Feb. 2023;13(4):2040. [Link]
[27] Ekinci S, Izci D, Jabari M, Bajaj M, Rubanenko O. Design of a novel and robust 2-DOF PIDA controller based on enzyme action optimizer for ball position
regulation in magnetic levitation systems. Sci Rep Aug. 2025;15(1):29360. [Link]
[28] Skliar M. Process dynamics and control. 2nd [Link]. 54, no. 11. John Wiley & Sons, Ltd; 2008. [Link]
[29] Luyben WL. Parallel Cascade Control. Ind Eng Chem Fundam Nov. 1973;12(4):463–7. [Link]
[30] Luyben WL. Chapter 2 : Examples of Mathematical Models for Chemical Processes. In: Examples of mathematical models of chemical engineering systems; 1990.
p. 28–102. Accessed: Dec. 14, 2025. [Online]. Available: [Link]
hl=hr&id=k7BTAAAAMAAJ.
[31] Lale T. GWO-WOA-AOA: Multistage hybrid metaheuristic optimization approach. In: In 2025 9th international symposium on innovative approaches in smart
technologies (ISAS). IEEE; Jun. 2025. p. 1–8. [Link]
[32] Yüksek G, Lale T. An improved thermal management of Lithium based batteries employing genetic algorithm optimization. Bull Pol Acad Sci Tech Sci Aug.
2024:151378. [Link]
[33] Huba M, Bistak P, Brieznik J, Vrancic D. Constrained series PI, PID and PIDA controller design inspired by Ziegler–Nichols. Power Electron Drives Jan. 2024;9
(1):331–46. [Link]
[34] Mintsa HA, Eny GE, Senouveau N, Nzué RMA. Optimal tuning PID controller gains from Ziegler-Nichols approach for an electrohydraulic servo system. J Eng
Res Reports Dec. 2023;25(11):158–66. [Link]
[35] Othman SM, et al. Position tracking performance with fine tune Ziegler-Nichols PID controller for electro-hydraulic actuator in aerospace vehicle model. J Phys
Conf Ser Nov. 2021;2107(1):012064. [Link]
[36] Ma D, Boussaada I, Chen J, Bonnet C, Niculescu S-I, Chen J. PID control design for first-order delay systems via MID pole placement: performance vs. robustness.
Automatica Mar. 2022;137:110102. [Link]
[37] Krohling RA, Rey JP. Design of optimal disturbance rejection PID controllers using genetic algorithms. IEEE Trans Evol Comput Feb. 2001;5(1):78–82. https://
[Link]/10.1109/4235.910467.
[38] Cao Jun-Yi, Liang Jin, Cao Bing-Gang. Optimization of fractional order PID controllers based on genetic algorithms. In: In 2005 international conference on
machine learning and cybernetics. vol. 9. IEEE; 2005. p. 5686–9. [Link]
[39] Saad M, Jamaluddin H. I. D. -, I. and Control, and undefined 2012, “Implementation of PID controller tuning using differential evolution and genetic
algorithms”. Int J Innov Comp Inform Control 2012;8(11):7761–79.
[40] Coelho LDS, Mariani VC. Firefly algorithm approach based on chaotic Tinkerbell map applied to multivariable PID controller tuning. Comput Math Appl Oct.
2012;64(8):2371–82. [Link]
[41] Mirjalili S, Mirjalili SM, Lewis A. Grey wolf optimizer. Adv Eng Softw Mar. 2014;69:46–61. [Link]
[42] Mirjalili S, Mirjalili SM, Saremi S, Mirjalili S. Whale optimization algorithm: theory, literature review, and application in designing photonic crystal filters. Stud
Comput Intell 2020;811:219–38. [Link]
[43] Jin Q, Qi L, Jiang B, Wang Q. Novel improved cuckoo search for PID controller design. Trans Inst Meas Control Jul. 2015;37(6):721–31. [Link]
10.1177/0142331214544211.
[44] Hung MH, Shu LS, Ho SJ, Hwang SF, Ho SY. A novel intelligent multiobjective simulated annealing algorithm for designing robust PID controllers. IEEE Trans
Syst Man Cybern A Syst Hum Mar. 2008;38(2):319–30. [Link]
[45] de Moura Oliveira PB, Solteiro Pires EJ, Novais P. Design of Posicast PID control systems using a gravitational search algorithm. Neurocomputing Nov. 2015;
167:18–23. [Link]
[46] Lee CH, Chang FK. Fractional-order PID controller optimization via improved electromagnetism-like algorithm. Expert Syst Appl Dec. 2010;37(12):8871–8.
[Link]
[47] Zhao W, Shi T, Wang L, Cao Q, Zhang H. An adaptive hybrid atom search optimization with particle swarm optimization and its application to optimal no-load
PID design of hydro-turbine governor. J Comput Des Eng Sep. 2021;8(5):1204–33. [Link]
[48] Dey S, Banerjee S, Dey J. Design of an improved robust fractional order PID controller for magnetic levitation system based on atom search optimization.
Sadhanā 2022;47(4):188. Sep. 2022, [Link]
[49] Chatterjee S, Mukherjee V. PID controller for automatic voltage regulator using teaching–learning based optimization technique. Int J Electr Power Energy Syst
May 2016;77:418–29. [Link]
[50] Yüksek G. A novel delay-affected two degree of freedom PID controller using physics-inspired optimization for robust control applications. Sci Rep Dec. 2025.
[Link]
[51] Gheisarnejad M. An effective hybrid harmony search and cuckoo optimization algorithm based fuzzy PID controller for load frequency control. Appl Soft
Comput Apr. 2018;65:121–38. [Link]
[52] Shabani H, Vahidi B, Ebrahimpour M. A robust PID controller based on imperialist competitive algorithm for load-frequency control of power systems. ISA Trans
Jan. 2013;52(1):88–95. [Link]
[53] Soukkou A, Khellaf A, Leulmi S, Boudeghdegh K. Optimal control of a CSTR process. Braz J Chem Eng 2008;25(4):799–812. [Link]
66322008000400017.
[54] Fan X, et al. Optimization of PID parameters based on improved particle-swarm-optimization. In: Proceedings - 2013 International Conference on Information
Science and Cloud Computing Companion, ISCC-C 2013; Dec. 2014. p. 393–7. [Link]
[55] Gaing ZL. A particle swarm optimization approach for optimum design of PID controller in AVR system. IEEE Trans Energy Convers Jun. 2004;19(2):384–91.
[Link]

25
S. Ekinci et al. Chaos, Solitons and Fractals 209 (2026) 118383

[56] Mistry K, Jani P, Tilva V. Comparative analysis of evolutionary algorithms for a CSTR system. In: In 2020 international conference on power, instrumentation,
control and computing (PICC). IEEE; Dec. 2020. p. 1–6. [Link]
[57] Begum KG. Coot bird optimization algorithm for the temperature control of continuous stirred tank reactor process. Asia Pac J Chem Eng Jan. 2023;18(1):
e2787. [Link]
[58] Der Chang W. Nonlinear CSTR control system design using an artificial bee colony algorithm. Simul Model Pract Theory Feb. 2013;31:1–9. [Link]
10.1016/[Link].2012.11.002.
[59] Ferreira JR, Sena AP, de S. Coutinho JP, Estumano DC, Macêdo EN. Fluid dynamics characterization of stirred-tank reactors via approximate Bayesian
computational (ABC) for parameter estimation and model selection. Numer Heat Transfer A Aug. 2024;85(16):2579–96. [Link]
10407782.2023.2226820.
[60] Alamdar Ravari M, Yaghoobi M. Optimum design of fractional order pid controller using chaotic firefly algorithms for a control CSTR system. Asian J Control
Sep. 2019;21(5):2245–55. [Link]
[61] Sharma A, Kar MK, Goud H. A novel modified grey wolf optimization tuned PID controller with fractional filter for the CSTR system. Heat Mass Transf 2025;61
(4):36. Apr. 2025, [Link]
[62] K. Mistry, P. Jani, and V. Tilva, “Comparative analysis of evolutionary algorithms for a CSTR System,” Proceedings of 2020 IEEE international conference on
power, instrumentation, control and computing, PICC 2020, Dec. 2020, doi:[Link]
[63] Haroon HEE, Huang C. BAT optimized SMC for continuous stirred tank reactor. IEEE Information Technology and Mechatronics Engineering Conference, ITOEC
2025;2025:1386–91. [Link]
[64] Gupta P, Kumar V, Rana KPS, Mishra P. Comparative study of some optimization techniques applied to Jacketed CSTR control. In: 2015 4th international
conference on reliability, Infocom technologies and optimization: trends and future directions, ICRITO 2015; Dec. 2015. [Link]
ICRITO.2015.7359368.
[65] Hong L, Mo Y, Bao D, Gong R. Chaos elite Harris hawk optimization algorithm to solve chemical dynamic optimization problems. IEEE Access 2022;10:
65833–53. [Link]
[66] Ajlouni N, Osyavas A, Ajlouni F, Almassri A, Takaoglu F, Takaoglu M. Enhancing PID control robustness in CSTRs: a hybrid approach to tuning under external
disturbances with GA, PSO, and machine learning. Neural Comput & Applic 2025;37(18):12153–77. Mar. 2025, [Link]
[67] Mistry K, Jani P, Tilva V. Comparative analysis of evolutionary algorithms for a CSTR System. In: Proceedings of 2020 IEEE international conference on power,
instrumentation, control and computing, PICC 2020; Dec. 2020. [Link]
[68] Li M, Chen H, Wang X, Zhong N, Lu S. An improved particle swarm optimization algorithm with adaptive inertia weights. Int J Inf Technol Decis Mak May 2019;
18(3):833–66. [Link]
[69] Khanduja N, Bhushan B, Mishra S. Control of CSTR using firefly and hybrid firefly-biogeography based optimization (BBFFO) algorithm. J Inf Optim Sci Aug.
2020;41(6):1443–52. [Link]
[70] v AR, D P, M T. A novel FOPID2FF2 controller optimized by a new Opposition based learning (OBL)-Black Widow Optimization (BWO) algorithm: application to
CSTR. Oct. 06, 2022. [Link]
[71] Kumavat M, Thale S. Analysis of CSTR temperature control with PID, MPC & Hybrid MPC-PID controller. ITM Web of Conferences 2022;44:01001. [Link]
org/10.1051/ITMCONF/20224401001.
[72] Cloutier JR, Stansbery DT. Control of a continuously stirred tank reactor using an asymmetric solution of the state-dependent Riccati equation. IEEE Conf
Control Appl Proc 1999;2:893–8. [Link]
[73] Braniff A, Tian Y. Reinforcement learning-based control via Y-wise Affine Neural Networks (YANNs). Accessed: Dec. 12, 2025. [Online]. Available: https://
[Link]/pdf/2508.16474v1; Aug. 2025.
[74] Sabura Banu U, Uma G. ANFIS based sensor fault detection for continuous stirred tank reactor. Appl Soft Comput Mar. 2011;11(2):2618–24. [Link]
10.1016/[Link].2010.10.009.
[75] Ahmad I. Advances in machine learning for monitoring, control, and optimization of temperature of reactors. Sep. 20, 2023. [Link]
preprints202309.1318.v1.
[76] Ekinci S, et al. Temperature control of nonlinear continuous stirred tank reactors using an enhanced nature-inspired optimizer and fractional-order controller.
Biomimetics Feb. 2026;11(2):153. [Link]
[77] Turkeri C, Ekinci S, Yüksek G, Li D. A modified enzyme action optimizer-based FOPID controller for temperature regulation of a nonlinear continuous stirred
tank reactor. Fractal Fraction Dec. 2025;9(12):811. [Link]
[78] Beyer WH. “Handbook of mathematical science,” Handbook of Mathematical Science. Jan. 2018. [Link]
[79] Zwillinger D, editor. CRC Standard Mathematical Tables and Formulae. Chapman and Hall/CRC; 2002. [Link]
[80] Wang L, et al. Tianji’s horse racing optimization (THRO): a new metaheuristic inspired by ancient wisdom and its engineering optimization applications. Artif
Intell Rev 2025;58(9):282. Jun. 2025, [Link]
[81] Wang J, Wang W, Hu X, Qiu L, Zang H. Black-winged kite algorithm: a nature-inspired meta-heuristic for solving benchmark functions and engineering
problems. Artif Intell Rev Mar. 2024;57(4):98. [Link]
[82] Zhao W, Wang L, Zhang Z. Atom search optimization and its application to solve a hydrogeologic parameter estimation problem. Knowl-Based Syst Jan. 2019;
163:283–304. [Link]
[83] Mirjalili S, Lewis A. The whale optimization algorithm. Adv Eng Softw May 2016;95:51–67. [Link]
[84] Srinivas M, Patnaik LM. Genetic algorithms: a survey. Computer (Long Beach Calif) Jun. 1994;27(6):17–26. [Link]
[85] O’Dwyer A. Handbook of PI and PID controller tuning rules. World Scientific Publishing Co; 2009. [Link]
[86] J. AK. PID controllers: theory, design, and tuning. Int Soc Measure Control 1995. [Link]
[87] Isaksson AJ, Graebe SF. Derivative filter is an integral part of PID design. IEE Proc Control Theory Appl Jan. 2002;149(1):41–5. [Link]
CTA:20020111.
[88] Bingi K, Ibrahim R, Karsiti MN, Hassan SM, Harindran VR. A comparative study of 2DOF PID and 2DOF fractional order PID controllers on a class of unstable
systems. Arch Control Sci Dec. 2018;28(4):635–82. [Link]
[89] Gorez R. New design relations for 2-DOF PID-like control systems. Automatica May 2003;39(5):901–8. [Link]
[90] Podlubny I. Fractional-order systems and fractional-order controllers. Slovak Academy of Sciences 1994;12(3):1–18.
[91] Monje C, Chen Y, Vinagre B, Xue D, Feliu V. Automatic control of a Hydraulic Canal. In: Advances in industrial control. Springer International Publishing; 2010.
p. 301–52. [Link] vol. no. 9781849963343.

26