This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

Deep Learning-based Model Predictive Control

for Resonant Power Converters
S. Lucia, Member, IEEE, D. Navarro, B. Karg, H. Sarnago, Senior Member, IEEE,
and O. Lucía, Senior Member, IEEE.

Abstract- Resonant power converters offer improved levels of

efficiency and power density. In order to implement such
systems, advanced control techniques are required to take the
most of the power converter. In this context, model predictive
control arises as a powerful tool that is able to consider
nonlinearities and constraints, but it requires the solution of
complex optimization problems or strong simplifying
assumptions that hinder its application in real situations.
Motivated by recent theoretical advances in the field of deep
learning, this paper proposes to learn, offline, the optimal
control policy defined by a complex model predictive
formulation using deep neural networks so that the online use of
the learned controller requires only the evaluation of a neural Fig. 1. Schematic description of Model Predictive Control. A mathematical
network. The obtained learned controller can be executed very model is used to predict the future states of a dynamic system and compute
rapidly on embedded hardware. We show the potential of the a sequence of optimal control inputs by solving an optimization problem.
presented approach on a Hardware-in-the-Loop setup of an
FPGA-controlled resonant power converter.
optimal control inputs by solving online a numerical
optimization problem that minimizes a chosen control goal
Keywords- Model predictive control, resonant power conversion, (see Fig. 1).
induction heating, deep learning, home appliances. MPC has been applied to power electronics converters [5,
17], specially to control electrical machines and drives.
I. INTRODUCTION However, among the many applications presented in literature
Since the appearing of resonant power conversion in the [18], little research has been published concerning resonant
70s [1-3], this technology has significantly evolved to provide power converters.
high performance and cost-effective power conversion in Because of the computational complexity of MPC, most
many different applications. Nowadays, the combination of research and applications of MPC for power electronics has
this technology with advances in digital control and focused on strongly simplified formulations. The most
semiconductor technology [4] has enabled the design of commonly used [17], is the Finite Control Set (FCS)
power electronic systems with a superior performance grade. approach, which chooses a control action between a discrete
The design of such systems also requires advanced control amount of possible switching strategies by testing all
techniques to take the most of these power converters [5]. possibilities. However, considering long horizons or variable
This includes also the development of modern digital control switching frequencies is more challenging. Other approaches
architectures [6], based either on FPGA [7, 8] or such as Explicit MPC [17] implement MPC as a look up table
microprocessor/DSP, that are enhanced by new development and are suitable for small linear systems. While there exist
techniques such as high level synthesis [9-11] and automatic improvements for both the FCS and advanced Explicit MPC
code generation [12], enabling the design of more advanced methods, it is still very challenging to design a controller that
and optimized controller implementations. can consider simultaneously nonlinearities, long horizons,
In recent years, a number of new control techniques have varying switching frequencies, time-varying constraints and
arisen including sliding-mode control [13] and model-free can be easily implemented. The motivation of this work is to
control [14, 15] applied to MIMO systems. In this context, propose an approach that can achieve all these goals.
model predictive control (MPC) has also arisen as an Instead of starting from strong simplifications, in this
effective tool to control complex systems [16] even when fast paper we design a rigorous nonlinear MPC problem.
sampling times are required. MPC uses a mathematical model Motivated by the latest advances in the field of deep learning
that describes the system dynamics to predict the future [19], [20], we use a deep neural network to directly learn the
trajectories of the system states and to compute a sequence of MPC solution. The online implementation of the controller is
thus reduced to the evaluation of a neural network, which
Manuscript received July 23, 2019; revised October 14, 2019; accepted January 22, requires less resources than explicit MPC and can cope with
2020. Copyright © 2020 IEEE. Personal use of this material is permitted. However,
permission to use this material for any other purposes must be obtained from the IEEE nonlinear systems and long horizons, as recently shown in
by sending a request to pubs-permissions@[Link]. [21, 22]. The use of machine learning to approximate MPC
This work was partly supported by the Spanish MINECO under Project TEC2016-
78358-R, by the Spanish MICINN and AEI under Project RTC-2017-5965-6, co-funded
by EU through FEDER program, by the DGA-FSE, and by DGA under Project
LMP106-18.
O. Lucia, D. Navarro and H. Sarnago are with the Department of Electronic
Engineering and Communications, University of Zaragoza, SPAIN (phone:
+34976761000, e-mail: olucia@[Link]). S. Lucia and B. Karg are with the Laboratory
of Internet of Things for Smart Buildings, Technische Universität Berlin, Germany (e-
mail: [Link]@[Link]).

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

Pan Vitroceramic
glass
SH

io RL Lr
Inductor Vs

Power Cr
electronics SL vo vc
Control

Fig. 3. Half-bridge series resonant converter.

Fig. 2. Induction heating home appliance.
appliances are a relevant example where cookers (Fig. 2)
controllers has also been recently proposed in [23], [24], [25], ranging from simple burners to complex multiload systems
[26]. With respect to [23] and [25], our approach considers [30, 31] are being developed.
nonlinear systems. In contrast to [24] and [26], we use deep Resonant power conversion plays a key role in such
neural networks to achieve an improved accuracy and we systems to create an alternating voltage to supply the
include a detailed discussion on data generation. Further induction coil, typically in the 20 kHz to 100 kHz frequency
novel contributions that were not considered in previous range. Usually, domestic IH systems implement the series-
results include an FPGA implementation, as well as resonant half-bridge converter (Fig. 3) due to its good balance
Hardware-in-the-loop (HIL) results and the consideration of between cost and performance. The half-bridge series
the tracking performance of the proposed controller under resonant inverter is composed of two switching devices, SH
uncertainty of model parameters. and SL, typically implemented using IGTBs, and the resonant
The aim of this paper is to propose a new approach to tank. The resonant tank is composed of the equivalent
design advanced MPC controllers to control resonant power impedance of the coil-pot system, RL, Lr, and the resonant
converters. More specifically, the proposed MPC scheme will capacitor Cr. The former parameters depend on the IH load
be applied to the control of a resonant inverter for induction selected by the user and its magnetic coupling, whereas the
heating (IH) applications. This application has a wide range latter can be considered constant. Thus, the differential
of operating conditions which severely affect the resonant equation system that describes the dynamics of the resonant
converter tuning, opening a wide field for control tank is defined by the current in the inductor, io,
optimization using MPC. In the past, few research studies
dio (t ) 1
Cr 
have dealt with IH applications. As far as the authors know, RL io (t )  Lr  io (t )dt  vo (t ). (1)
only in [27] a model predictive controller is proposed, with dt
limited insight into the controller itself. Further contributions The modulation strategies typically used to control the
of this paper include a novel time transformation that enables output power include square wave modulation and
a simple MPC formulation to explicitly handle important asymmetrical duty cycle [32]. The former is based on
constraints such as zero-voltage switching. This paper modifying the switching frequency and the latter is based on
extends the results presented in [28] by including novel modifying both, duty cycle D and switching frequency fsw.
contributions as the approximation of the MPC control law The maximum output power is obtained at the resonant
based on deep learning and incorporating FPGA
frequency, defined as fo  1 2 Lr Cr . In order to achieve
implementation and HIL simulations. This implementation is
intended for a final cost-effective ASIC implementation for zero voltage switching to reduce switching losses, fsw and D
domestic induction heating systems. are increased to reduce the output power. If the required
The remainder of this paper is organized as follows. output power is too low, typically pulse density modulation
Section II reviews IH technology and Section III covers (PDM) is applied in order to limit the maximum fsw.
model predictive control and its application to the half-bridge The control strategy of the series resonant inverter for IH
series resonant converter. Section IV and V summarizes the applications is usually constrained by several aspects inherent
main neural network and FPGA implementation results, to the application, being the most important the following:
respectively. Finally, Section VI summarizes the main Load variability: The IH load, geometry and coupling, and
conclusions of this paper. required output power vary within a wide range, changing the
resonant tank and, consequently, making the design of the
II. RESONANT POWER CONVERSION FOR IH SYSTEMS controller a challenging task [33]. Classical PI controllers
Induction heating [29] has become the leading heating must be tuned to respond to fixed plant characteristics,
technology due to its benefits in terms of efficiency, safety whereas MPC controllers can be constrained to actuate in the
and performance. The applications of IH have been extended desired manner under a wider range of operating conditions.
in recent years to cover a wide range of industrial, domestic Soft-switching: In order to optimize the converter
and medical applications. Among these, domestic IH efficiency and safe operation, it is essential to guarantee soft-

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

switching in the whole operational range [32], i.e. achieve TABLE I

SUMMARY OF THE MODEL STATE AND INPUT VARIABLES
ZVS conditions. This is specially challenging when using
linear controllers, where some external constraints are Variables Description
difficult to be taken into account and may lead to unexpected
State variable output current: Partly defines the second-
performance and/or instability. MPC controllers allow to io order resonant circuit. It is proportional to the created
introduce these conditions, based on specific circuit states at magnetic field and, therefore, the induced heating.
the switching moment, as an additional constraint, whereas it
is not feasible to use this approach with classical linear State variable resonant capacitor voltage: Completes the
vc
definition of the second-order resonant circuit.
controllers.
Multi-load control: Induction heating systems are usually Input variable switching frequency of the inverter devices:
composed by several loads that may operate simultaneously. In a series resonant circuit, it is used to control the output
fsw
power and is, typically, equal or higher than the resonant
Under these circumstances, it is essential to guarantee that frequency to achieve soft-switching conditions.
they operate synchronously, and no acoustic noise is Input variable duty ratio of the half-bridge inverter
generated due to switching frequencies intermodulation. In devices: is the activation ratio of the switching devices
this context, MPC offers powerful tools for control during a switching period. In the proposed half-bridge
D
configuration, the output voltage/power is maximum at D
optimization where a wide range of constrains, state space = 0.5, and they decrease when D increases/decreases with
variables and control actions are present, optimizing the symmetric behavior.
overall switching performance.
The central challenge, especially for applications in power
Special operating conditions: Under some conditions, the
electronics, is that the implementation of MPC requires the
controller must be adapted due to excessive temperature, low
online solution of the following optimization problem at each
electromagnetic coupling, maximum power consumed by
sampling time:
several loads or limiting flicker emissions. All these specific
N 1
issues are usually addressed by saturating or introducing
exceptions in classical controllers. However, this can lead to
minimize
uk
 J (x ,u
k 0
k k )
unpredictable operation or put at risk the converter safety and subject to xk 1  f ( xk , uk ), x0  xˆ (3)
performance. MPC provides a powerful control environment
g ( xk , uk )  0
where all these restrictions can be taken into account without
degrading the controller performance and reliability. where, g describes general constraints and x̂ is the current
All these constrains severely limit the control possibilities state of the system, which needs to be measured or estimated
of the converters and possess a significant challenge to design every time a new control should be computed. The
the control strategy. Moreover, as it has been discussed, all performance metric, usually called cost function, is denoted
these constraints are difficult to be implemented using by J and N is the prediction horizon.
classical linear control schemes, increasing the complexity of In model predictive control, from the computed sequence
the controller and limiting its scalability. For these reasons, of optimal control inputs (u0, …, uN-1) only the first element
MPC is explored as an effective control scheme to provide (u0) is applied to the system and the same optimization
improved control and versatility for resonant power problem is solved at the next step with the new state of the
converters applied to IH. system. For applications that require sampling times in the
range of microseconds, solving (2) in real-time can be very
III. MODEL PREDICTIVE CONTROL
challenging despite of the recent progress in hardware,
A. General formulation algorithms and tailored implementations [34] that have
enabled the application of MPC to fast systems.
Model predictive control uses the mathematical model of a In the field of power electronics, most implementations of
dynamic system to predict its future behavior and compute a MPC [17] are based on a discretization of all control
sequence of optimal control inputs that satisfies given possibilities (FCS-MPC). A different possibility is the use of
constraints and optimizes a desired performance measure. explicit MPC [35] which precomputes the solution of the
The model of a dynamic system can be written in a optimization problem for all possible states, so that the online
discrete-time setting as: implementation reduces to a search algorithm.
xk 1  f ( xk , uk ), (2) Explicit MPC has been very successful for small linear
systems but its implementation is more complex in the
where xk   n denotes the state of the system at time k and nonlinear case and for long horizons, because the amount of
uk   m the control inputs. One of the main strengths of memory needed to store the controller grows exponentially
MPC is that it can consider nonlinear systems, general with the prediction horizon and the number of constraints. A
constraints, as well as control tasks different than classical comparison of both continuous control set (CSS) and finite
setpoint tracking. control set (FCS) MPC is presented in [36].
We propose in this work the use of deep learning to learn

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

the optimal control policy, as done in approximate explicit integer variables. However, if combined with long prediction
MPC [37], by means of a deep neural network (DNN) that horizons and time-varying constraints, it results in very
can be easily implemented on an FPGA. This approach opens complex mixed-integer optimization problems. We propose
the door for the implementation of complex nonlinear MPC to perform a double time transformation that transforms the
schemes with constraints and long horizons in fast systems, mixed-integer problem into a standard nonlinear but
which can be implemented in FPGAs with a small need of continous optimization problem.
digital resources that could be massively deployed in most We propose a time-varying transformation that maps the
power converters. We do not consider theoretical guarantees original time t to the new scaled time units  as follows:
of stability or recursive feasibility as done e.g. in [38].  f sw,i
 t if t  (ti , ti  Di / f sw,i ]
B. MPC for Resonant Inverters  Di
  (6)
The resonant tank described in (1) can be modeled using  f sw,i t if t  (ti  Di / f sw,i , ti 1 ]
the state-space paradigm. Table I summarizes the main state 1  Di
and input variables used in this model. Considering two states
The main advantage of this transformation is that the
x = [io, vc], the system dynamics can be written as:
switching in the new time units occurs exactly at   1, 2,3,
dio 1 We make use of this property to discretize the differential
 (vo  RL io  vc ),
dt Lr equations (3) in the new time units with a discretization time
(4) of  s  1 units. The model that is used in the MPC
dvc 1
 io . formulation can be written as:
dt Cr
The available control inputs are the switching frequency  Dk [on ]
f f ( xk , uk ) if k  0, 2, 4,..., N
and the duty cycle u = [fsw, D], which determine the signal vo.  sw,k
One of the main challenges for the control of the half-bridge xk[mpc]
1  (7)
series resonant tank is the switched nature of vo which makes  Dk  1 f [off ] ( x , u ) if k  1,3,5,..., N  1.
 fsw,k k k
the dynamic system (3) a hybrid system. The signal vo can be 
written as: Where f [on ] and f [off ] denote discretized version of the
V if t  (ti , ti  D / fsw ] differential equations in (3) when vo  Vs and vo  0
vo   s (5)
0 if t  (ti  D / fsw , ti 1 ] respectively.
Solving general MPC problems of switched systems is a This reformulation greatly simplifies the MPC formulation,
very difficult problem that leads to difficult optimization as it is possible to consider long prediction horizons N
problems, especially if non-linearities and time-dependent without the use of integer variables. It is also possible to use
constraints are considered, as it occurs for the computation of varying switching frequencies and duty cycles for each
the average power or the consideration of ZVS constraints. switching interval.
Relying on the finite control set (FCS) MPC technique is D. Average Power Computation
difficult in this case because we need to include time-varying
constraints and make use of varying frequencies and duty To discretize the time-transformed version of the dynamics
cycles to improve efficiency. Instead, we propose to consider (7) to be used in the prediction of the MPC, we use
the full nonlinear model, with long horizons if necessary, and orthogonal collocation on finite elements [39]. Each control
with flexible frequency and duty cycle in an efficient interval is divided into finite elements, on which the state
formulation that facilitates the offline solution of a large trajectory is parametrized using Lagrange polynomials. This
number of MPC problems. To enable the real-time discretization scheme is often used in nonlinear model
implementation of the scheme, we propose to generate large predictive control as it provides a superior accuracy compared
amounts of data by solving rigorous nonlinear MPC problems to simpler Euler discretization schemes. Using orthogonal
and learn the control law using deep learning as will be collocation, the states of the system can be computed at any
explained in Section IV. point in time and are directly available at the collocation
The proposed nonlinear model predictive controller for points, which are new optimization variables.
induction heating has three key elements: a time The average power, calculated for each switching interval,
transformation, the average power computation, and the can be computed as:
1/ fsw
consideration of ZVS constraints. We explain these Po  f sw  io (t )vo (t )dt (8)
components in the remainder of the section 0

To achieve a tractable online computation of the average

C. Time transformation
power, we use the value of the states at the collocation points
The switching frequency and the duty cycle determine the within each control interval. Because of the time
switching instants t k of the system defined in (3). The direct transformation previously described, each switching interval
consideration of the switching instants can be modeled using is composed of two control intervals (one step k is used for

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

the ON semi-cycle and one for the OFF semi-cycle). The

50
average power at each control interval can be thus calculated

i o [A]
0
as: -50
400
 ncol

 f sw , k  io , k vo , k ( k   k ),

v c [V]
[ i 1]
[i ] [i ] [i ]
if k  0, 2,, N 200
Po , k  i 1 (9) 0
P , if k  1,3,, N  1, 300
 o , k 1

f sw [kHz] v o [V]
200
100
where ncol denotes the number of collocation points in each 0
100
control interval and the io[i,]k denotes the value of the current at 50
control step k for collocation point i. The integral is
1
approximated by a simple quadrature in which  k[i ] denotes

D [-]
0.5
the time of collocation point i in the control interval k. 0
This implies that the average power can be predicted in

P o [W]
2000
advance, even during transient behavior, regardless of the
0
inputs used (frequency, duty cycle, etc …) provided that the 0 1 2 3
model is correct. Several techniques can be used to obtain Time [s] # 10
-4

good estimations of RL and Lr, see [33, 40] for an overview of Fig. 4. Simulation results for the state variables, switching signal, control
possible methods. We also show in Section V that it is inputs and power. MPC controller when both fsw and D are available as
control inputs. The red dotted line denotes the desired power signal.
possible to adapt the MPC formulation to deal with
uncertainty in RL and Lr.
equations and power computation are included in constraints
E. ZVS Constraints (10b), and (10c, 10d) denote the input constraints. The ZVS
Another strength of the proposed time transformation is constraints are enforced by (10e, 10f). The fact that the inputs
that enforcing zero-voltage-switching reduces to imposing cannot be changed within a switching interval is enforced by
static constraints at the end of each control interval. In the (10g, 10h).
series resonant converter under study, ZVS depends on the The optimization problem is nonconvex because of the
operating conditions and can be achieved by ensuring the power computation and because of the control-dependent
correct output current sign during the turn-on transition [3, time transformation.
41]. These constraints can be easily incorporated in any MPC The results presented here use the primal-dual interior
framework. It is straightforward to introduce any additional point method described in [42] and implemented in IPOPT to
constraints that might be required by a specific application, as solve all the optimization problems. Within the interior point
the ones described in Section II. algorithm, an exact Hessian is used. All the derivative
The optimization problem that should be solved after each information is computed using automatic differentiation via
switching cycle, based on the current measurements of the CasADi [43] and the MPC loop is implemented using the
toolbox do-mpc [44]. The real system is simulated with the
states io , vo , and the power setpoint Podes is:
numerical integrator Sundials [45]. We use a prediction
N horizon of N  10 steps, which equals 5 switching intervals.
minimize
fsw ,k , Dk
 (P
k 0
o,k  Podes ) 2   fsw, k (a) The length of the horizon is reasonable as it is the typical
number of switching cycles that the system needs to achieve a
subject to model in (7) (b)
new steady state. The model is discretized with orthogonal
30  f sw, k  100 kHz, (c ) collocation based on Lagrange polynomials of degree 2, with
0.2  Dk  0.8, (d ) (10) 100 finite elements in each control interval to achieve an
accurate average power computation. The tuning parameter
io , k  0, if k  1,3,5,..., N  1, (e )
that weights the importance in reducing the switching
io , k  0, if k  0, 2, 4,..., N , (f) frequency is chosen as α =5e-8 in order to balance the
Dk  Dk 1 , if k  1,3,5,..., N  1, (g) relative contributions in the cost function.
f sw, k  f sw , k 1 , if k  1,3,5,..., N  1. (h) The simulation results obtained by solving Problem (10) at
each switching interval are shown in Fig. 4. The NMPC
The tuning parameter α in the cost function (10a) weights controller is switched on after 5 cycles and three different
the relative importance between using a small frequency and steps in the desired power are performed after 5 cycles each
tracking the desired power. It is usually desired to choose the (500 W, 3000 W and 1000 W). The model parameters are
minimum possible frequency, as this maximizes the Vs=230 V, Lr=19ꞏ10-6 H, RL=2.9 Ω, and Cr=1440 nF. It can be
efficiency [32]. Note that the model used in the MPC problem seen that the different power setpoints (red dashed line in
is a discretized version of the model described in (4) using the bottom plot of Fig. 4) can be tracked accurately while
orthogonal collocation on finite elements and including the using both control inputs and respecting the ZSV constraints
double time transformation described in (6). The discretized continuously, also during the transient behavior.

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

OFFLINE CONTROLLER
CONTROLLER DESIGN DEPLOYMENT
Training scenarios Performance evaluation Optimization & Verification

Offline NMPC DNN C‐coded VHDL

solver training DNN code
io, vo, Po fsw, D W, b High Level Synthesis

Fig. 6. Schematic representation of the controller design and deployment of

the proposed deep learning-based NMPC solution.

(a)

Fig. 7. Representation of a deep neural network with L hidden layers with

M neurons each. The input layer and the output layer have nin and nout layers
respectively.

Intel Core i7 with 16 GB RAM). To obtain an implementable

controller, we propose to approximate the solution of a
rigorous NMPC formulation using deep neural networks
instead of oversimplifying the problem formulation.
(b)
Fig. 5. Comparison performance of the proposed MPC controller with PI
implementations: (a) dynamic performance and (b) waveforms detail
where D PI do not achieve ZVS conditions. IV. IMPLEMENTATION USING DEEP NEURAL NETWORKS
To illustrate the advantages of the MPC controller, Fig. 5 The embedded implementation of described controller is
(a) shows a comparison performance with classical PI usually a challenging task due to real-time requirements and
controllers. It is important to note that solving problems for hardware restrictions. In this paper, an effective deep neural
variable-plant MIMO systems would require otherwise using network implementation taking advantage of fast and parallel
gain-schedulling techniques [46]. In this comparison, we computing in FPGA will be proposed [6] (Fig. 6).
consider a PI controller that uses the error in the power The approximation of the solution of MPC control laws
tracking to adapt the switching frequency fsw and another PI has been usually studied under the headline of approximate
controller that uses the tracking error of the power to adapt explicit MPC [37]. While neural networks had been
the duty cycle D. The tuning of the PI controllers is previously proposed to learn the mapping between states and
performed via simulations and is chosen so that the fastest control inputs [46], using deep neural networks for this
possible performance without overshoot in the delivered purpose has been studied only very recently (see e.g. [21,
power is obtained. To facilitate the comparison between 22]).
approaches, the x-axis represents switching cycles instead of Using deep neural networks (with several hidden layers),
time. It can be seen that the MPC approach is significantly instead of classical shallow ones (with only one hidden layer)
faster than the PI controller for the switching frequency fsw. has been one of the key developments that have enabled the
While the PI controller for the duty cycle D can obtain a practical successes of deep learning in previous years [19].
similar performance when compared to MPC, it cannot The latest theoretical advances [20] have shown that deep
incorporate constraints, resulting in violations of ZVS as it neural networks have greater expressivity than shallow
can be seen in Fig. 5 (b). Note that the MPC controller also networks, leading to better approximation capabilities of
enables the incorporation of any other constraint such that complex functions. This idea has been recently exploited in
avoidance of acoustic noise due to frequency coupling [22] to propose deep neural networks to approximate the
between inductors. In same manner, several inductors could function defined by the solution of the MPC problem. In that
be controller simultaneously with the proposed MPC while case, the inputs for the neural network are the current state of
solving problems for variable-plant MIMO systems would the system (and potentially also other parameters), and the
require otherwise using gain-scheduling techniques [45] that outputs of the network are the optimal control inputs. Since
still cannot incorporate constraints. the control strategy is defined via a neural network, our
The solution of each NMPC problems takes in average proposed approach could take advantage of future computing
around 500 ms in a laptop computer (Macbook Pro, 3.5 GHz

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

50
where u * denotes the optimal solution obtained when solving the
i o [A]
0
-50 MPC Problem (10) which is a function of the current state and the
400 current setpoint. The inputs of the neural network  are also the
vc [V]
200
0 current states and the current power setpoint and N s is the number
300 of samples that are used for the training.
f sw [kHz] vo [V]

200
100
0
The training is done with Tensorflow [47] via Keras [48],
100 using the optimizer ADAM [49], which is a modification of
50 stochastic gradient descent and all computations are
performed with 32-bit floating point number representation.
1
Fig. 8 shows the results of the deep-learning based
D [-]

0.5
0
controller for the same control task as the one shown in Fig.
4. It can be seen that the desired power setpoint is
P o [W]

2000 successfully tracked and the ZVS constraints are not violated.
0
0 1 2 3 Both trajectories, for the exact solution of the NMPC and the
Time [s] # 10 -4 proposed deep learning-based NMPC are very similar.
Fig. 8. Simulation results for the state variables, switching signal, control Table II presents a systematic evaluation of the
inputs and power when both fsw and D are available control inputs performance of the proposed controller. We run 100 different
computed via deep learning approximation of the MPC control law. The
red dotted line denotes the desired power signal. simulations (similar to the ones shown in Fig. 4) for randomly
TABLE II
chosen power setpoints. The different power setpoints are
COMPARISON OF THE AVERAGE TRACKING PERFORMANCE INCLUDING sampled between 500 W and 3000 W following a uniform
TRANSIENT BEHAVIOR AND ZVS CONSTRAINT VIOLATIONS FOR 100 distribution. As shown in Table II, the proposed controller
DIFFERENT RANDOM SETPOINT CHANGES WITH R AND L EXACTLY KNOWN achieves almost the same tracking performance than the exact
Average tracking error Average ZVS violation solution of the NMPC problem and does not result in any
Controller
[W/cycle] [% of cycles]
violation of the ZVS constraints.
Exact NMPC 5.6939 0 These promising results suggest that using the proposed
Deep learning- method, it is possible to approximate very accurately a high-
6.2551 0
based NMPC performance NMPC solution using a deep neural network that
hardware tailored for the fast and efficient computation of can be easily deployed on embedded hardware, as we show in
machine learning tasks. the next sections.
A deep neural network (Fig. 7) is a sequence of layers of V. DATA GENERATION AND ROBUSTNESS ANALYSIS
neurons that determines a function  with inputs z   nin The number of data pairs that are used to train the neural
that can be defined as: network and the way that such data pairs are generated has a
 ( z )  f L 1  g L  f L    g1  f1 ( z ), (11) significant influence on the approximation quality that can be
achieved.
where L is the number of hidden layers and M the number of
We chose a neural network with L=5 hidden layers and
neurons in each layer and ° denotes function composition.
M=10 neurons per layer for all cases. We did not observe
Each hidden layer consists of an affine function
significant improvements for larger networks. The training
f l ( l 1 )  Wl  l 1  bl , (12) data is composed of several input-output pairs. The input
where l 1   M is the output of the previous layer and includes the power setpoint Podes , as well as the current
 0  z. The matrices Wl and the vectors bl are called the measurements of io and vc and the output includes the optimal
inputs fsw and D as illustrated in Fig. 7. For a given input, the
weights and biases at layer l and their value is determined by output can be automatically generated by solving the MPC
training the neural network with known input-output pairs. problem (10).
The second component of the neural network is the nonlinear The input for which the corresponding outputs are
function g l () , which is called activation function. Usual calculated can be sampled randomly from the allowed state
choices for g l () include the rectifier linear unit (ReLU), the space in which the system can be operated. We sample this
sigmoidal function or the tanh function. space using a uniform distribution for each one of the three
inputs with the ranges Podes  [0,3000]W, io  [ 150,150]A
Neural networks are trained offline, which means that the and vc  [2000, 2000]V .
optimal values of the weights W [ l ] and biases b[l ] are If the optimization problem for the chosen initial state is
computed by minimizing the mean squared error: unfeasible, the data point is discarded. The training error for
1 Ns * different amounts of data points can be seen in Fig. 9.
minimize
Wl ,bl

N s s 1
(u (xˆs , Podes )   ( xˆs , Podes ))2 , (13) An alternative approach to generate training data is to use
an initial condition from the same allowed space but then
simulate the closed-loop behavior using the MPC controller

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

Fig. 9. Training error along the training iterations (epochs) for the same
neural network structure when different amount of training data is used and
different methods for the generation of the training data.

for a given amount of simulation steps. In this way, we Fig. 10. Simulation results for the state variables, switching signal,
generate training samples around optimal trajectories, which control inputs and power when both fsw and D are available control inputs
is where the system is usually operated and we avoid the use computed via deep learning approximation of the MPC control law. The
of states which might never be reached by the system. Using red dotted line denotes the desired power signal. There is a 15% error for
the model parameters RL and Lr .
this kind of data generation, the training error can be reduced
by an order of magnitude (see Fig. 9) and that is why this within a 15 % error interval with respect to the ones that were
method is used to generate the data for the considered used for training. In this figure, ZVS operation can be also
example. As expected, increasing the amount of data also observed, verified by the correct current sign during the turn-
leads to an improvement of the training error in any case. The on transition. A summary of the results can be seen in Table
validation error on previously unknown trajectories has a III. The tracking error is again very similar to the case when
similar behaviour and it is omitted for brevity. Even if the the optimization problem is solved at each sampling time.
network is trained on optimal trajectories, the obtained Although ZVS violations are slightly smaller for the deep
controller can be robust to different model parameters as we learning-based controller, the difference is negligible and it in
illustrate in the following. both cases only minor ZVS violations occur. If strict ZVS is
Ensuring the robustness of the controller with respect to desired for each cycle, robust NMPC techniques could be
model parameters is important since they are usually not used.
exactly known or can vary rapidly. Achieving zero steady VI. EMBEDDED IMPLEMENTATION RESULTS
state error in NMPC under the presence of uncertainty can be
often achieved by updating the desired power setpoint with All the embedded results presented in this section take
the current power using a simple linear update rule. That is, at advantage of the time transformation presented in (7) and
each sampling time the desired power used in the cost which enables the fast solution of the MPC problems that
function of the MPC is updated following: generate the training data for the neural networks, as
described in Section IV and Section V. The main motivation
Podes  Podes  K ( Podes, orig  Podes, meas ), (14)
to approximate the solution of MPC via a neural network is
where Podes, orig is the original desired power setpoint, Podes, meas that it enables a very simple embedded implementation of a
is the currently measured power, and K = 0.8 is a tuning complex controller, as explained in the remainder of the
parameter that can be chosen to balance the speed of the section.
corrective actions. If the correction term (14) is not used, the A. FPGA Implementation
formulation is the same as the one in (10), where Podes  Podes,orig
The proposed MPC controller using deep neural networks
is constant and Podes,meas is not used. An additional advantage of is implemented using FPGA technology to take advantage of
the proposed deep learning-based MPC controller is that fast and parallel processing. Moreover, recent advances in
exactly the same correction rule can be applied. In this case, FPGA technology [6] have led to cost-effective devices with
the updated power setpoint is chosen as input to the neural an increasing amount of digital resources, enabling high-
network. As a result, steady state accuracy is achieved even performance and competitive implementations.
under the presence of model errors. Fig. 10 shows the Neural networks can be implemented with low precision
simulation results for the proposed approximated NMPC arithmetic without degrading performance [50]. In order to
controller with a 15 % error in the model parameters RL and quantize and normalize coefficients, in this work we use
Lr. The correction strategy is activated at the middle of the Ristretto [51], an automated Neural Networks approximation
simulation time and it can be seen that the correction leads to tool which condenses 32-bit floating point networks. Ristretto
no tracking error despite of the uncertain parameters. We also is an extension of Caffe [52] and allows to test, train and fine-
performed 100 simulations for random values of RL and Lr tune networks with limited numerical precision.

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

In order to explore the design space and obtain an DIGITAL HW

optimized implementation, automatic code generation using LAYER i, m
high level synthesis (HLS) [9, 53] has been applied. The neurons
implementation has been optimized to achieve the desired W i,1 Ni,1 W i,2 N i,2 W i,n N i,m
timing performance while minimizing the required digital N i+1,1 N i,1 Ni+1,1 Ni,2 N i+1,1 Ni,m

HW requirements. As it is shown in Fig. 11, each neuron in N i+1,1 N i,1 Ni+1,1 Ni,2 N i+1,1 Ni,m

each layer is calculated sequentially in order to minimize HW

resources while achieving the required timing performance BIAS

(see Table IV). By using the following HLS directive, the

j=1
ReLU j=2

maximum number of multipliers is limited to 10: j=n

HLS ALLOCATION instances=mul limit=10 operation

N i+1,1 N i+1,2 N i+1,m

After that, the same digital HW is used to calculate the SEQUENTIAL EXECUTION (INNER NEURON LOOP)

next layer, improving the final implementation. This is done SEQUENTIAL EXECUTION (OUTER LAYER LOOP)
by using directive:
LAYER i+1, m neurons
HLS ALLOCATION instances=fcc limit=1 function
Fig. 11. Optimized arithmetic FPGA implementation.
Finally, in order to optimize usage, each layer coefficients
are stored in the same memory using dual-port memories. By TABLE III
COMPARISON OF THE AVERAGE TRACKING PERFORMANCE INCLUDING
doing that, only 5 memories are required to feed data to the TRANSIENT BEHAVIOR AND ZVS CONSTRAINT VIOLATIONS FOR 100
10 multipliers. This is implemented by using: DIFFERENT RANDOM SETPOINT CHANGES WITH 15 % ERROR IN R AND L
Average tracking error Average ZVS violation
HLS array_map variable=weights_1 instance=weights horizontal Controller
[W/cycle] [% of cycles]
HLS array_map variable=weights_2 instance=weights horizontal
Exact NMPC 10.845 0.3922
Table IV shows the performance for floating point and 16 Deep learning-
12.243 0.0784
bits implementation. The fixed-point version reaches the based NMPC
target latency of 1 µs while keeping the digital HW resources TABLE IV
to low levels. The error with respect to the floating-point FPGA IMPLEMENTATIONS SUMMARY
implementation is less than 5% in the worst case. Item 16-bit implementation Floating-point implementation
B.A. HIL results
Latency 1 µs 2.1 µs
Simulation and verification of induction heating systems LUTs 1382 12851
are challenging due to the high variety of operating FFs 888 7396
conditions including output power levels and different
DSPs 10 50
resonant tanks determined by the pot materials, geometry and
temperature. These parameters are usually randomly chosen in the FPGA following the implementations detailed at the
by the user and, consequently, cannot be predicted. beginning of this Section.
Furthermore, new materials and geometries usually appear in Fig. 13 shows an example of a post-layout simulation with
the market to be used with deployed appliances that should be 16-bit arithmetic. In this figure, the current through the
operative for years. In this context, HIL has proven to be a resonant tank, the target output power and the controller error
very effective tool to simulate and test proposed controllers are represented. As it can be seen, the controller works as
and control strategies under a wide variety of conditions [10, expected, achieving the desired output power while keeping
33, 54] in a safer and faster way and, consequently, it will be ZVS constraints, which can be verified checking the proper
used in this paper. current sign during the switching times. These results prove
The proposed HIL platform used for domestic induction the feasibility of the proposed controller, as well as its FPGA-
heating verification is depicted in Fig. 12. It consists on an implementation using deep neural networks and high-level
FPGA-based architecture that runs both ad-hoc digital synthesis.
hardware and software running in the dual-core ARM To further validate the robustness of the proposed
processor available. Linux OS is used to generate controller, we performed several HiL experiments using the
automatically the required test-benches as well as to perform deep learning-based NMPC with the correction term (14)
results verifications and communications with a user which are shown in Table V. In particular, we perform nine
interface. A plant model of the series resonant inverter is also different experiments where three different steps in the
implemented, which operates in real time, taking as inputs the desired power output Podes ,orig are performed. The results
control variables and the operating parameters, i.e. resonant
tank, and generating as outputs the state variables. The reported in Table V include the steady state accuracy (after 10
proposed deep learning-based MPC controller is implemented switching cycles) of the controller with 16-bit arithmetic as
well as the number of ZVS violations for different values of

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

the parameters RL and Lr. In all cases, the proposed controller real time advanced control of power converters.
achieves a very small steady state error for different values of The proposed scheme has been detailed, and the controller
the parameters thanks to the correction term (14). performance has been successfully validated in simulations
and in a hardware-in-the-loop setup. The proposed controller
VII. CONCLUSIONS enables to take into account complex control constraints for
Resonant power conversion enables the implementation of IH systems as well as to obtain improve dynamics in a
power converters with superior performance, efficiency and variable-plant MIMO system that would require otherwise
power density. Among the multiple applications, induction more complex and difficult to tune control techniques.
heating systems are a relevant example which outperforms Furthermore, the proposed control architecture enables the
other technologies due to its superior efficiency and implementation of a complex MPC controller in a simple
performance. However, the accurate control of the resonant manner in state-of-the-art ASIC implementations with
converter under a wide variety of constraints of different significant benefits in terms of control constraints and
nature still remains a challenge. This paper has proposed how dynamic performance. As a conclusion, the proposed
high-performance NMPC solutions can be easily and approach opens the door to the application of complex
accurately approximated using deep neural networks. The NMPC-schemes for future higher-performance higher-
resulting neural networks can be used in a High-Level complexity IH [Link] work will study the scalability
Synthesis framework to obtain FPGA designs that enable the properties of the proposed approach and further comparisons
with other approximate explicit MPC methods as done in the
linear case in [22].
REFERENCES
[1] R. Oruganti and F. C. Lee, "Resonant power processors, Part II-
Methods of control," IEEE Transactions on Industry Applications, vol.
21, no. 6, pp. 1461-1471, November/December 1985.
[2] R. Oruganti and F. C. Lee, "Resonant power processors, Part I---State
plane analysis," IEEE Transactions on Industry Applications, vol. 21,
TABLE V
STEADY STATE TRACKING ERRORS FOR THE HIL IMPLEMENTATION WITH
DIFFERENT SETPOINTS AND DIFFERENT PARAMETER VALUES

Steady state error (mW) for different setpoints

R error L error
Average ZVS
1000 W 2000 W 3000 W
violation [%]
0% 0% 41.74 1.554 2.747 0
0% -15% 6.903 2.149 0.699 0
0% +15% 96.22 448.1 737.7 0
-15% 0% 10.39 2.023 97.46 0
-15% -15% 8.632 3.964 5.97 0
-15% +15% 223.0 449.3 54.19 0
+15% 0% 48.46 22.68 125.1 0
+15% -15% 2.234 3.282 3.693 0
+15% +15% 49.78 116.9 108.3 0
Fig. 12. Hardware-in-the-loop testbench.

Fig. 13. Post-layout simulation using 16-bit arithmetic.

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

no. 6, pp. 1453-1460, November/December 1985. [24] M. Hertneck et al., "Learning an Approximate Model Predictive
[3] M. K. Kazimierczuk and D. Czarkowski, Resonant Power Converters. Controller With Guarantees," IEEE Control Systems Letters, vol. 2, no.
New York: John Willey & Sons, 2011. 3, pp. 543-548, 2018.
[4] X. She, A. Q. Huang, O. Lucia, and B. Ozpineci, "Review of Silicon [25] J. Drgoňa, D. Picard, M. Kvasnica, and L. Helsen, "Approximate model
Carbide Power Devices and Their Applications," IEEE Transactions on predictive building control via machine learning," Applied Energy, vol.
Industrial Electronics, vol. 64, no. 10, pp. 8193-8205, Oct 2017. 218, pp. 199-216, 2018/05/15/ 2018.
[5] S. Kouro, M. A. Perez, J. Rodriguez, A. M. Llor, and H. A. Young, [26] B. M. Åkesson and H. T. Toivonen, "A neural network model
"Model Predictive Control: MPC's Role in the Evolution of Power predictive controller," Journal of Process Control, vol. 16, no. 9, pp.
Electronics," IEEE Industrial Electronics Magazine, vol. 9, no. 4, pp. 8- 937-946, 2006/10/01/ 2006.
21, 2015. [27] P. Kranprakon, A. Sangswang, S. Naetiladdanon, and E.
[6] O. Lucía, E. Monmasson, D. Navarro, L. A. Barragán, I. Urriza, and J. Mujjalinwimut, "A model predictive control of an LLC resonant
I. Artigas, "Modern control architectures and implementation," in inverter for tin melting application," in IECON 2017 - 43rd Annual
Control of Power Electronic Converters and Systems, vol. 2, Frede Conference of the IEEE Industrial Electronics Society, 2017, pp. 3773-
Blaabjerg, Ed. 1 ed. USA: Elsevier, 2018. 3778.
[7] E. Monmasson, L. Idkhajine, M. N. Cirstea, I. Bahri, A. Tisan, and M. [28] S. Lucía, D. Navarro, H. Sarnago, and O. Lucía, "Model predictive
W. Naouar, "FPGAs in industrial control applications," IEEE Trans. control for resonant power converters applied to induction heating," in
Ind. Informat., vol. 7, no. 2, pp. 224-243, May 2011. IEEE International Symposium on Industrial Electronics, 2018, vol. 1,
[8] E. Monmasson, L. Idkhajine, and M. W. Naouar, "FPGA-based pp. 246-251.
controllers," IEEE Trans. Ind. Electron. Magaz., vol. 5, no. 1, pp. 14- [29] O. Lucía, P. Maussion, E. Dede, and J. M. Burdío, "Induction heating
26, 2011. technology and its applications: Past developments, current technology,
[9] S. Lucia, D. Navarro, O. Lucia, P. Zometa, and R. Findeisen, and future challenges," IEEE Transactions on Industrial Electronics,
"Optimized FPGA Implementation of Model Predictive Control for vol. 61, no. 5, pp. 2509-2520, May 2014.
Embedded Systems Using High Level Synthesis Tool," IEEE [30] O. Lucía, J. Acero, C. Carretero, and J. M. Burdío, "Induction heating
Transactions on Industrial Informatics, vol. 14, no. 1, pp. 137-145, appliances: Towards more flexible cooking surfaces," IEEE Industrial
2017. Electronics Magazine, vol. 7, no. 3, pp. 35-47, September 2013.
[10] O. Jimenez, O. Lucia, I. Urriza Parroque, L. A. Barragan, D. Navarro, [31] H. Sarnago, P. Guillén, J. M. Burdío, and O. Lucía, "Multiple-Output
and V. Dinavahi, "Implementation of an FPGA-based on-line hardware- Soft-Switching Resonant Inverter Architecture for Flexible Induction
in-the-loop emulator using high-level synthesis tools for resonant power Heating Appliances," IEEE Access, vol. 7, pp. 157046-157056, 2019.
converters applied to induction heating appliances," IEEE Transactions [32] O. Lucía, J. M. Burdío, I. Millán, J. Acero, and L. A. Barragán,
on Industrial Electronics, vol. 62, no. 4, pp. 2206-2214, April 2015. "Efficiency oriented design of ZVS half-bridge series resonant inverter
[11] D. Navarro, O. Lucía, L. A. Barragán, I. Urriza, and O. Jiménez, "High- with variable frequency duty cycle control," IEEE Transactions on
level synthesis for accelerating the FPGA implementation of Power Electronics, vol. 25, no. 7, pp. 1671-1674, July 2010.
computationally-demanding control algorithms for power converters," [33] O. Jiménez, O. Lucía, I. Urriza, L. A. Barragán, and D. Navarro,
IEEE Transactions on Industrial Informatics, vol. 9, no. 3, pp. 1371- "Analysis and implementation of FPGA-based online parametric
1379, August 2013. identification algorithms for resonant power converters," IEEE
[12] B. Khusainov, E. C. Kerrigan, and G. A. Constantinides, "Automatic Transactions on Industrial Informatics, vol. 10, no. 2, pp. 1144-1153,
Software and Computing Hardware Codesign for Predictive Control," May 2014.
IEEE Transactions on Control Systems Technology, pp. 1-10, 2018. [34] B. Houska, H. J. Ferreau, and M. Diehl, "An auto-generated real-time
[13] G. Wang and Q. Xu, "Sliding Mode Control with Disturbance Rejection iteration algorithm for nonlinear MPC in the microsecond range,"
for Piezoelectric Nanopositioning Control," in 2018 Annual American Automatica, vol. 47(10), pp. 2279–-2285, 2011.
Control Conference (ACC), 2018, pp. 6144-6149. [35] A. Bemporad, [Link], V. Dua, and E. N. Pistikopoulos, "The explicit
[14] R. Roman, M. Radac, and R. Precup, "Multi-input–multi-output system linear quadratic regulator for constrained systems," Automatica, vol. 38,
experimental validation of model-free control and virtual reference pp. 3-20, 2002.
feedback tuning techniques," IET Control Theory & Applications, vol. [36] A. A. Ahmed, B. K. Koh, and Y. I. Lee, "A Comparison of Finite
10, no. 12, pp. 1395-1403, 2016. Control Set and Continuous Control Set Model Predictive Control
[15] A. Safaei and M. N. Mahyuddin, "Optimal model-free control for a Schemes for Speed Control of Induction Motors," IEEE Transactions
generic MIMO nonlinear system with application to autonomous on Industrial Informatics, vol. 14, no. 4, pp. 1334-1346, 2018.
mobile robots," International Journal of Adaptive Control and Signal [37] T. A. Johansen, "Approximate explicit receding horizon control of
Processing, vol. 32, no. 6, pp. 792-815, 2018. constrained nonlinear systems," Automatica, vol. 40, pp. 293-300, 2004.
[16] J. B. Rawlings and D. Q. Mayne, "Model Predictive Control Theory and [38] J. B. Rawlings and D. Q. Mayne, Model Predictive Control: Theory and
Design," 2009. Design. Nob Hill Pub., 2009.
[17] S. Vazquez, J. Rodriguez, M. Rivera, L. G. Franquelo, and M. [39] L. T. Biegler, "Nonlinear Programming: Concepts, Algorithms, and
Norambuena, "Model Predictive Control for Power Converters and Applications to Chemical Processes," 2010.
Drives: Advances and Trends," IEEE Transactions on Industrial [40] O. Jiménez, O. Lucía, L. A. Barragán, D. Navarro, J. I. Artigas, and I.
Electronics, vol. 64, no. 2, pp. 935-947, 2017. Urriza, "FPGA-based test-bench for resonant inverter load
[18] S. Vazquez et al., "Model Predictive Control: A Review of Its characterization," IEEE Transactions on Industrial Informatics, vol. 9,
Applications in Power Electronics," IEEE Industrial Electronics no. 3, pp. 1645-1654, August 2013.
Magazine, vol. 8, no. 1, pp. 16-31, 2014. [41] H. Sarnago, O. Lucía, A. Mediano, and J. M. Burdío, "Analytical model
[19] Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. of the half-bridge series resonant inverter for improved power
521, p. 436, 2015. conversion efficiency and performance," IEEE Transactions on Power
[20] B. Hanin, "Universal Function Approximation by Deep Neural Nets Electronics, vol. 30, no. 8, pp. 4128-4143, August 2015.
with Bounded Width and ReLU Activations," pp. 1-14, 2017. [42] A. Wächter and L. Biegler, "On the implementation of a Primal-Dual
[21] S. Lucia and B. Karg, "A deep learning-based approach to robust Interior Point Filter Line Search Algorithm for Large-Scale Nonlinear
nonlinear model predictive control," in Proc. of the 6th IFAC Programming," Mathematical Programming, vol. 106, pp. 25-57, 2006.
Conference on Nonlinear Model Predictive Control, ed, 2018, pp. 610- [43] J. Andersson, J. Åkesson, and M. Diehl, "{CasADi} -- {A} symbolic
615. package for automatic differentiation and optimal control," in Recent
[22] B. Karg and S. Lucia, "Efficient representation and approximation of Advances in Algorithmic Differentiation, ed, 2012, pp. 297-307.
model predictive control laws via deep learning," arXiv preprint [44] S. Lucia, A. Tatulea-Codrean, C. Schoppmeyer, and S. Engell, "Rapid
arXiv:1806.10644, 2018. Development of Modular and Sustainable Nonlinear Model Predictive
[23] S. Chen et al., "Approximating Explicit Model Predictive Control Using Control Solutions," Control Engineering Practice, vol. 60, pp. 51-62,
Constrained Neural Networks," in 2018 Annual American Control 2017.
Conference (ACC), 2018, pp. 1520-1527.

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2020.2969729, IEEE
Transactions on Industrial Informatics

IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS

[45] A. C. Hindmarsh et al., "{SUNDIALS}: {S}uite of Nonlinear and [51] P. Gysel, J. Pimentel, M. Motamedi, and S. Ghiasi, "Ristretto: A
Differential/Algebraic Equation Solvers," ACM Transactions on Framework for Empirical Study of Resource-Efficient Inference in
Mathematical Software,, vol. 31, pp. 363-396, 2005. Convolutional Neural Networks," IEEE Transactions on Neural
[46] T. Parisini and R. Zoppoli, "A Receding-Horizon Regulator for Networks and Learning Systems, pp. 1-6, 2018.
Nonlinear-Systems and a Neural Approximation," Automatica, vol. 31, [52] Y. Jia et al., "Caffe: Convolutional Architecture for Fast Feature
pp. 1443-1451, 1995. Embedding," ArXiv e-prints, Accessed on: June 01, 2014Available:
[47] M. Abadi et al., "TensorFlow : A System for Large-Scale Machine [Link]
Learning This paper is included in the Proceedings of the TensorFlow : [53] Xilinx, Introduction to high-level synthesis with Vivado HLS
A system for large-scale machine learning," 2016. Standalone (Xilinx). 2016.
[48] F. Chollet. (2015). Keras. Available: [Link] [54] O. Lucía, I. Urriza, L. A. Barragan, D. Navarro, O. Jiménez, and J. M.
[49] D. P. Kingma and J. L. Ba, "ADAM: A Method for Stochastic Burdío, "Real-time FPGA-based hardware-in-the-loop simulation test-
Optimization," pp. 1-15, 2015. bench applied to multiple output power converters," IEEE Transactions
[50] I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio, on Industry Applications, vol. 47, no. 2, pp. 853-860, March/April
"Quantized Neural Networks: Training Neural Networks with Low 2011.
Precision Weights and Activations," ArXiv e-prints, Accessed on:
September 01, 2016Available:
[Link]

Sergio Lucia (M’16) received the [Link]. degree Hector Sarnago (S’09 M'15 SM'19) received
in electrical engineering from the University of the [Link]. degree in Electrical Engineering
Zaragoza, Zaragoza, Spain, in 2010, and the Dr. and the Ph.D. degree in Power Electronics
Ing. degree in optimization and automatic from the University of Zaragoza, Spain, in
control from the Technical University of 2010 and 2013, respectively. Currently, he is
Dortmund, Dortmund, Germany, in 2014. He a senior post-doc researcher in the the
joined the Otto-von-Guericke Universitat Department of Electronic Engineering and
Magdeburg and visited the Massachusetts Communications at the University of
Institute of Technology as a Postdoctoral Fellow. Zaragoza, Spain. His main research interests
Since May 2017, he has been an Assistant Professor and Chair with include resonant converters and digital control for induction heating
the Laboratory of Internet of Things for Smart Buildings, applications.
Technische Universitat Berlin, Berlin, Germany, and with Einstein Dr. Sarnago is a member of the Aragon Institute for
Center Digital Future, Berlin. His research interests include Engineering Research (I3A).
decision-making under uncertainty, distributed control, as well as
the interplay between machine learning techniques and control
theory. Dr. Lucia is currently Associate Editor of the Journal of
Process Control. Óscar Lucía (S’04, M’11, SM’14) received
the [Link]. and Ph.D. degrees (with honors) in
Denis Navarro received the [Link]. degree in Electrical Engineering from the University of
Microelectronics from the University of Zaragoza, Spain, in 2006 and 2010,
Montpellier, France, and the Ph.D. degree from respectively.
the University of Zaragoza in 1987 and 1992, During 2006 and 2007 he held a
respectively. research internship at the Bosch and Siemens
Since September 1988, he has been with the Home Appliances Group. Since 2008, he has
Department of Electronic Engineering and been with the Department of Electronic
Communications at the University of Zaragoza, Engineering and Communications at the University of Zaragoza,
where he is a Professor. His current research Spain, where he is currently an Associate Professor. During part of
interests include CAD for VLSI, low power ASIC design, and 2009 and 2012, he was a visiting scholar at the Center of Power
modulation techniques for power converters. He is involved in the Electronics Systems (CPES), Virginia Tech. His main research
implementation of new applications of integrated circuits. In 1993 interests include resonant power conversion, wide-bandgap devices,
Dr. Navarro designed the first SPARC® microprocessor in Europe. and digital control, mainly applied to contactless energy transfer,
Dr. Navarro is a member of the Aragon Institute for Engineering induction heating, electric vehicles, and biomedical applications. In
Research (I3A). these topics, he has published more than 70 international journal
papers and 150 conference papers, and he has filed more than 40
Benjamin Karg was born in Burglengenfeld, international patents.
Germany, in 1992. He received the [Link]. Dr. Lucía is a Senior Member of the IEEE and an active member of
degree in mechanical engineering from the Power Electronics (PELS) and Industrial Electronics (IES)
Ostbayerische Technische Hochschule societies. He was a Guest Associate Editor of the IEEE Transactions
Regensburg, Regensburg, Germany, in 2015, on Industrial Electronics and the IEEE Journal of Emerging and
and his [Link]. degree in systems engineering Selected Topics in Power Electronics in 2013 and 2015,
and engineering cybernetics from Otto-von- respectively. Currently, he is an Associate Editor of the IEEE
Guericke Universität, Magdeburg, Saxony- Transactions on Industrial Electronics, IEEE Open Journal of
Anhalt, Germany, in 2017. He currently works as a research Industrial Electronics, and IEEE Transactions on Power Electronics.
assistant at the laboratory "Internet of Things for Smart Buildings", Dr. Lucía is a member of the Aragon Institute for Engineering
Technische Universität Berlin, Germany, to pursue his PhD. He is Research (I3A).
also member of the Einstein Center for Digital Future. His research
is focused on control engineering, artificial intelligence and edge
computing for IoT-enabled cyber-physical systems.

1551-3203 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See [Link] for more information.
Authorized licensed use limited to: University of Canberra. Downloaded on April 30,2020 at 13:01:54 UTC from IEEE Xplore. Restrictions apply.