Environmental Science and Pollution Research (2023) 30:84474–84490

[Link]

RESEARCH ARTICLE

Long‑lead streamflow forecasting using computational intelligence

methods while considering uncertainty issue
Mohammad Najafzadeh1 · Sedigheh Anvari2

Received: 17 November 2022 / Accepted: 9 June 2023 / Published online: 27 June 2023
© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2023

Abstract
While some robust artificial intelligence (AI) techniques such as Gene-Expression Programming (GEP), Model Tree (MT),
and Multivariate Adaptive Regression Spline (MARS) have been frequently employed in the field of water resources, docu-
ments aimed to explore their uncertainty levels are few and far between. Meanwhile, uncertainty determination of these AI
models in practical applications is highly important especially when we aimed to use the AI models for streamflow forecast
due to the repercussions of poorly managed water resources. With the aid of a global daily streamflow dataset, understanding
the uncertainty of GEP, MT, and MARS for forecasting streamflow of natural rivers was studied. The efficiency of uncer-
tainty analysis was quantified by two statistical indicators: 95% Percent Prediction Uncertainty (95%PPU) and R-factor.
The results demonstrated that MT had lower uncertainty (95%PPU=0.59 and R-factor=1.67) in comparison with MARS
(95%PPU=0.61 and R-factor=1.92) and GEP (95%PPU=0.64 and R-factor=2.03). Overall, although the confidence interval
bands of uncertainty for the AI models almost captured the mean streamflow measurements, wide bands of uncertainty were
obtained and consequently remarkable uncertainty in the calculation of monthly streamflow values was met.

Keywords Streamflow forecast · Uncertainty analysis · Artificial intelligence models · Statistical measures

Notations BF Basis Function

a0 Bias term of multivariate linear equation CC Correlation of Coefficient, R.
a1, a2, a3 Weighting coefficients of multivariate CM Conventional Model
linear equation d Average distance between the lower and
ACF Auto-Correlation Function upper 95PPU band
ANFIS Adaptive Neuro-Fuzzy Inference System EPR Evolutionary Polynomial Regression
AI Artifiucuial Intelligence FFNN Feed Forward Neural Network
AR Auto-Regression GA Genetic Algorithm
ARMA Auto-Regression Moving Average GEP Gene-Expression Programming
ARIMA Auto-Regression Integrated Moving GLUE Genetalized Likelihood Uncertainty
Average Estimation
GMDH Group Method of Data Handling
Responsible Editor: Marcus Schulz KNN K-Nearest Neighbor
KELM Kernel Extreme Learning Machine
* Mohammad Najafzadeh MARS Multivariate Adaptive Regression Spline
[Link]@[Link]; [Link]@[Link] MCS Monte-Carlo Simulation
Sedigheh Anvari MSE Mean Square Error
anvari.t@[Link]; [Link]@[Link] MT Model Tree
1
Department of Water Engineering, Faculty of Civil NBF Number of Basis Function
and Surveying Engineering, Graduate University P-factor The key element for quantifying the
of Advanced Technology, P.O. Box 76315117, Kerman, Iran uncertainty of AI models performance
2
Department of Ecology, Institute of Science and High PACF Partial Auto-Correlation Function
Technology and Environmental Sciences, Graduate PSO Particle Swarm Optimization
University of Advanced Technology, P.O. Box 76315117, Q Monthly or dialy streamflow discharge
Kerman, Iran

13
Vol:.(1234567890)
Environmental Science and Pollution Research (2023) 30:84474–84490 84475

RE Relative Error approaches is inextricably bound up with the completeness

R-factor The key elements for quantifying the of time series, yielding reliable forecasts. For instance, the
uncertainty of AI models performance ARMA model is an appropriate technique especially in the
RC Coefficient of Correlation case of short-term time series forecasting and it has the
RMSE Root Mean Square Error capability to assess the estimation accuracy while consider-
SF Streamflow ing the number of variables to a minimum. In accordance
SHDI Standardized Hydrological Drought with this, the performance of these traditional techniques
Index provided forecasts with a low accuracy level while treating
SOI Southern Oscillation Index with short-time series and incomplete time series (Shahabi
STD Standard Deviation and Azarpira 2021).
SVM Support Vector Machine With the rapid rise of development in Artificial Intel-
t Time ligence (AI) models, a wide range of newly advanced AI
t−1, t−2, t−3 Lag times of streamflow discharge techniques were applied to forecast the streamflow in rivers
TF Thomas-Fiering method and natural streams. For instance, Artificial Neural Network
WRM Water Resources Management (ANN), Adaptive Neuro-fuzzy Inference System (ANFIS),
xU Upper limit of 95PPU K-Nearest Neighbor (K-NN), Model Tree (MT), Group
xL Lower limit of 95PPU Method of Data Handling (GMDH), Multivariate Adap-
ψ0 Bias term of MARS model tive Regression Spline (MARS), Support Vector Machine
ψi Weigthing coefficients of MARS model (SVM), and Gene-Expression Programming (GEP) have
95PPU 95 Percent Prediction Uncertainty been successfully performed to enhance their effectiveness
in forecasting streamflow at various temporal scales (Marce'
et al. 2004; Kisi 2005; Noori et al. 2010; Noori et al. 2011;
Introduction El-Shafie and Noureldin 2011; Saghafian et al. 2013; Schnier
and Cai 2014; Anvari et al. 2014; Cheng et al. 2015; Yaseen
Streamflow forecasts are generally considered one of the et al. 2015; Yang et al. 2017; Shiri 2018; Niu et al. 2018;
most unavoidable issues which make significant contribu- Modaresi et al. 2018; Al-Sudani et al. 2019; Poul et al. 2019;
tion to the management of water resources projects such as Hussein and Khan, 2020; Noorbeh et al. 2020; Yaseen et al.
the operation of flood control reservoirs, design of hydraulic 2020; Shahabi and Azarpira 2021; Amininia and Saghebian
structures, surface water quality, and recharge of groundwa- 2021; Riahi-Madvar et al. 2021; Moosavi et al. 2022).
ter resources. In this way, more reliable streamflow forecasts From the above-mentioned AI models, many AI models
are key element of managing water resources with vari- (i.e., MT, GEP, and MARS) could provide reliable math-
ous climatic circumstances, particularly in arid and semi- ematical equations to forecast streamflow when compared
arid regions (e.g., Saghafian et al. 2013; He et al. 2014; to other AI models such as ANN, ANFIS, and SVM models,
Tikhamarine et al. 2020; Moghaddasi et al. 2022). Various working as a black box. In the case of streamflow forecast-
flow components contribute to streamflow generation which ing, the performance of formulation-based AI models dem-
have become the cornerstone of hydrological investigations onstrated high potential in the interpretation of the results
for more than two decades. Direct runoff and snowmelt are than those AI models performed on the basis of black box
the main source of streamflow generation. These elements principles. The availability of meteorological and hydrologi-
play a key role in causing variations in streamflow with the cal variables causes uncertainties in the streamflow forecast
seasonal and annual scales. These spatial and temporal vari- due to the fact that these variables vary temporally and spa-
ations make streamflow forecasting a much more challeng- tially. Although formidable efforts were made to reach recent
ing issue (i.e., Setegn et al. 2010; Saghafian et al. 2013; improvements in an increase in the precision level of com-
Narsimlu et al. 2015). putational power and diminishing volume of computations,
Since the late 1980s, a broad range of studies have been paying meticulous attention to the uncertainty in relation
conducted to forecast streamflow encompassing different with hydrological events is an inevitable issue. Overall, the
lead times and forecast techniques. Quite a few Conventional quantification of uncertainty is dependent on understanding
Models (CMs) (e.g., Auto-Regression [AR], Auto-Regres- prior knowledge of uncertainty. Over the past decades, the
sion Moving Average [ARMA], and AR Integrated Moving usability of statistical models along with hydrological mod-
Average [ARIMA]) were evaluated to demonstrate precision els and AI models has discovered that there are four sources
level of streamflow forecast for a large number of investiga- of uncertainty occurring in deterministic flow modeling. The
tions (Zealand et al. 1999; Coulibaly et al. 2000; Bazart- first source is related to random or systematic error values
seren et al. 2003; Nayak et al. 2004; Kisi 2005; Aqil et al. in the numerical model input variables as boundary or ini-
2007). The high potential of performance of conventional tial conditions are used. The second source of uncertainty

13
84476 Environmental Science and Pollution Research (2023) 30:84474–84490

occurs in the recorded output variables recruited to compute Thomas-Fiering methodology; (ii) to perform the MT, GEP,
simulation precision. Third, uncertainties are due to values and MARS for the modeling streamflow using observed/
of sub-optimal parameter. Ultimately, uncertainties are due deterministic values of streamflow at Armand hydrometric
to structure of biased or incomplete numerical models (Yu station, located just upstream of Karoon-4 Dam site in south-
and Tseng 1996; Diomede et al. 2006; Wanger et al. 2007; western Iran; (iii) to evaluate the performance of AI models
Bensoussan and Farhi, 2010; Noori et al. 2011; Yan et al. for 100 replicates by using statistical measures and report
2012; Bayazit, 2015; Amininia and Saghebian 2021). In this the most accurate evaluations; and (iv) to employ an implicit
way, the uncertainty issues considering their various sources approach of MCS procedure to investigate the uncertainty
requires to be efficiently applied for hydrological investiga- analysis of the AI models by quantity-based indices and
tions in order to understand practical viewpoints related to quality-based indices. Figure 1 depicts workflow of the pre-
environments. sent investigation.
In previous studies of streamflow forecast, the AI models
and conventional techniques which sometimes suffer from
various cases (e.g., limited range of datasets system, exist- Literature review: forecasting models
ing irregular time series datasets, and non-continuous time regarding uncertainty
series) have been employed to yield the relationships in the
limit state. The development of these methods was carried Over the past decades, considering uncertainty concepts in
out for reliable datasets with a certain range of input-output the forecasting of hydrological studies have played a key role
variables under various field conditions such as climate in the design of robust hydrological models. In this way, a
changes, severe floods, and lack of amenities for flow rivers broad range of studies were conducted to quantify the uncer-
measurement. In these investigations, the uncertainty and tainty analysis of AI models for forecasting models. Some
randomness of datasets were excluded to understand the literatures are presented herein:
performance of AI models. To eradicate this limitation, the In Wanger et al.’s (2007) research, over many years of
uncertainty analysis is regarded as an effective solution to investigation, an uncertainty system has been enhanced in
study the impact of random variability of input-output vari- order to forecast flash-flood in the semi-arid southwestern
ables on the accuracy assessment of AI models. Focusing USA. They applied the Generalized Likelihood Uncer-
on the uncertainty research of AI models for forecasting the tainty Estimation (GLUE) methodology to conduct uncer-
streamflow has not been completely studied (i.e., Noori et al. tainty analysis. With reference to the setting parameters of
2011; Dehghani et al. 2014; Amininia and Saghebian 2021); GLUE approach, results of uncertainty analysis in the set-
therefore, the current studies still require further develop- ting parameters of the hillslope model indicated a higher
ment to find what extent uncertainty analysis affects the influence on the approximations than the uncertainties in
performance of AI models and to understand the effects of the geometric parameters of channel, at least for relatively
setting parameters and general structure of AI models on the small area basins. Montanari and Grossi (2008) proposed
computational errors of predictions/forecasts. a methodology for the uncertainty quantification of hydro-
The AI models, namely GEP, MT, and MARS models, logical forecasts. They could find sources of uncertainties
are all examples of symbolic regression techniques. These related to computational errors in forecasts. In addition, Di
techniques are particularly suitable when the goal is to Baldassarre and Montanari (2009) quantified the uncertainty
obtain a mathematical formula that describes the relation- analysis that is likely due to streamflow river observations
ship between the input variables and the output variable in a for a 330-km reach of the Po River, Italy. From their results,
form that is easily interpretable. In contrast, ANN, ANFIS, it was found that computational errors in streamflow values
and SVM models are examples of non-symbolic machine are not absolutely negligible.
learning algorithms. These AI models are powerful tools Pham (2011) used GLUE method to perform the param-
for building predictive models, but their black box nature eterization uncertainty for three basins in France. He consid-
makes it difficult to interpret the AI models prediction. In ered three various sources of uncertainty: precipitation data,
some cases, this lack of interpretability may be acceptable, hydrological initial conditions (streamflow time series data),
especially when the goal is to achieve high accuracy in pre- and model parameters. From his investigation, the results indi-
diction tasks. cated that input precipitation uncertainty did not have signifi-
To the authors’ knowledge, the reported uncertainty cant effects on the forecast output. He recommended consid-
investigations of discharge forecast on the performance ering the uncertainty analysis of forecast precipitation, initial
of MT, GEP, and MARS models are few and far between. condition, and hydrological model parameters in flood fore-
Accordingly, the contemporary contributions of the pre- cast. Pablo et al. (2012) evaluated the effectiveness of a flood
sent study are (i) to generate monthly streamflow data- forecasting framework in a sparsely monitored basin, located
series in order to carry out stochastical analysis by using in southern Chile. Additionally, they considered simulation of

13
Environmental Science and Pollution Research (2023) 30:84474–84490 84477

Fig. 1  Flowchart of the adopted

methodology

geospatial data, assimilation analysis, and numerically mod- obtain the indices of the upper and lower forecast members
eled meteorological variables with verification of uncertainty of an ensemble for a case study in Bavaria, Germany. They
quantification in explicit way. They demonstrated streamflow concluded that the proposed methodologies were compu-
uncertainty more efficiently during flood occurrences. In tationally efficient owing to the fact that they significantly
Saghafian et al.’s (2013) investigation, the effect of meteoro- plummeted the number of members of the ensembles needed
logical variables with spatial distribution as well as Southern to generate a flood flow forecast with high predictive skill.
Oscillation Index (SOI) on the monthly flow forecasts was Furthermore, Amininia and Saghebian (2021) applied two
studied by soft computing models. They concluded that usa- robust AI models, namely kernel extreme learning machine
bility of the SOI as an additional variable slightly enhanced (KELM) and MARS, to forecast the monthly stream flow of
the accuracy level of ANN and ANFIS models performance. the Mississippi River, USA, over a 70-year period beginning
Dehghani et al. (2014) analyzed drought forecasts by con- in 1950. They found the permissible degree of uncertainty for
verting the time series of monthly streamflow into the Stand- the most efficiently utilized model.
ardized Hydrological Drought Index (SHDI). With the aid
of the SHDI time series, they simulated the drought event by
using feed-forward neural network (FFNN). From their study, Material and methods
the performance of FFNN indicated that all simulated values
of streamflow were placed within the 95% confidence inter- Study area
vals. Steenbergen and Willems (2015) investigated three main
uncertainty sources coming from model, recorded rainfall, and The great Karoon river basin, with a basin drainage area of
the initial conditions for hydrological flood forecast in the Bel- 71,980 ­km2, is located in southwestern Iran between 49° 30′
gian case study of Rivierbeek, Belgium. They could reduce to 52° eastern longitude and 30° 30′ to 32° northern latitude
the model uncertainty, as the leading factor of uncertainty, (Fig. 2). Karoon River is introduced as the longest natural
by using a stepwise-based calibration methodology. Leandro stream which has 867-km meandering length. The Karoon
et al. (2019) developed three forecast methods in order to River streams include over one-fifth of the country’s surface

13
84478 Environmental Science and Pollution Research (2023) 30:84474–84490

Fig. 2  Location of the Great Karoon basin and the case study of sub-basin in southwestern Iran

waters. Generally, Karoon River originates from Zagros moun- scattered dry farming, bare land, urban districts), types of basin
tain ranges and passes through Khuzestan plain. Then, this soils (i.e., brown steppe soils, desert soils, lithosols soils, non-
river is emptied into the Persian Gulf. There are quite a few saline alluvial soils), and minor and major streamflow networks.
urban areas which the Karoon River across them and addi- This study forecasts streamflow forecasts at Armand hydro-
tionally Ahvaz city is considered a megacity that is the center metric station, located just upstream of Karoon IV dam site in
of Khuzestan province. More importantly, there are various southwestern Iran, by GEP, MT, and MARS models. Figure 4
major industries (i.e., steel, agriculture, and oil) that are highly illustrates variations of the daily and monthly discharge (Q) val-
dependent on the Karoon River. Hence, the Karoon River inevi- ues versus time for the Karoon river at Armand station which
tably plays a prominent role in driving economy forward in the covers a 28-year period beginning at 1981. As seen in Fig. 4a,
country. Moreover, large concrete and earth dams were built the highest amount of streamflow is related to the months of
on the Karoon River which plays a key role in managing sur- May (478.86 m ­ 3/s), March (403.39 m­ 3/s), and April (356 m­ 3/s),
face waters during flood events, generating hydropower energy, which is due to the rains of the previous months as well as runoff
and providing agricultural water management. The major vol- from snow melting. On the other hand, the lowest amount of
ume of streamflow of the Karoon River naturally originates streamflow belongs to September (21.21 m ­ 3/s), October (21.33
from snowmelt-driven peak flows in March and April whereas ­m3/s), and March (23.05356 ­m3/s). Additionally, Fig. 4b illus-
theKaroon River experiences low volume of streamflows in trates that there is a significant negative trend in which the peak
September and October (Afkhami et al. 2007; Saghafian et al. of monthly streamflows in the spring season plummets. In addi-
2013; Riahi-Madvar et al. 2021). Furthermore, Fig. 3 illustrates tion, the monthly streamflow in the summer has a low volume
distribution of land use conditions (e.g., irrigated farming, rock, of discharge whereas low streamflow is on the rise in autumn.

13
Environmental Science and Pollution Research (2023) 30:84474–84490 84479

Fig. 3  Landscape characteristics of Karoon-4 River basin: (a) land use, (b) soil type, and (c) stream network

Figure 4b demonstrates that the downward trend is more likely resources management (WRM). Generally, the existing uncertainty
due to the presence of dams and climate change along with of WRM issue comes from three major sources, namely structure,
decrease in precipitation. The main statistical characteristics of input variables, and setting parameters of numerical models. In the
the Aramand hydrometric station are also presented in Table 1. case of hydrological models, the uncertainty in meteorological and
hydrological variables (inputs of models such as rainfall, tempera-
Implementation of soft computing models ture, and streamflow) and setting parameters is likely influenced
by the computational errors in the measurement of inputs and
Stochastic data generation and uncertainty analysis errors related to the non-uniqueness of setting parameters. Hence,
the accurate assessment of the forecast uncertainty is inevitably
Precision evaluation of the prediction/forecast uncertainty is known influenced by the methodologies utilized to deal with uncertainty
as a fundamental tool for dealing with design problems in water sources (Abbaspour 2008; Zhang et al. 2009; Anvari et al. 2022).

13
84480 Environmental Science and Pollution Research (2023) 30:84474–84490

Fig. 4  Time series of stream-

flow in the Armand hydromet-
ric station for the period of
1981–2008: (a) daily and (b)
monthly scales

Table 1  Statistical characteristics of the streamflow in the Aramand In the present study, to consider the uncertainty related
hydrometric station to streamflow forecasts, 100 replicates of these monthly
Statistics Daily streamflow Montly streamflow predictors have been produced with aid of the Thomas-
Fiering (TF) model (1962). The TF methodology is gener-
Number of data points 10220 336 ally performed as the stochastic technique that has been
Standard deviation [­ m3/s] 95.00 79.08 widely utilized to forecast variables in hydrological inves-
Skewness (Pearson) 3.931 1.88
tigations (e.g., McMahon and Adeloye 2005; Cui et al.
Minimum value [­ m3/s] 20.10 21.21
2016; Soundharajan et al. 2016). The technique used to
25% quantile [­ m3/s] 43.50 43.93
generate the 100 datasets can contribute to the uncertainty
50% quantile (median) ­[m3/s] 61.60 65.67
reflected in the AI models predictions. The uncertainty in
Mean ­[m3/s] 96.75 96.60
the AI models predictions arises from various sources such
75% quantile [­ m3/s] 114.00 116.15
as the variability in the data, the assumptions made in the
Maximum value ­[m3/s] 1489.91 12.00
model, and the limitations of the modeling technique. In
Coefficient of variance 9025.46 6253.65
the case of generating datasets, the method used to create
these datasets can influence the variability in the data and
thus contribute to the uncertainty in the model prediction.
The usability of MCS had efficient performance to pro- For instance, if the datasets are generated using a random
duce knowledge extraction of the best predictors while sampling technique, different random samples can lead
considering the uncertainties of meteorological and to different datasets with different statistical properties,
hydrological variables in streamflow forecasts (Adeloye, which can affect the model’s predictions.
2012; Kroese et al. 2014; Rubinstein and Kroese 2016). This methodology can be efficiently employed to gen-
The MCS is generally performed based on the concept of erate larger datasets because of limited availability of
randomly repeated samplings (or synthetically produced observed datasets (or poorly monitored datasets). Addi-
sets) for an uncertain variable. Statistical characteristics tionally, TF model can provide more diverse datasets
of random datasets are approximately the same as those in order to simulate extreme events when hydrological
obtained from field data measurements (Talebizadeh et al. observations have low variability. On the other hand, TF
2010; Dehghani et al. 2014). is capable of improving quality of observed streamflow

13
Environmental Science and Pollution Research (2023) 30:84474–84490 84481

values (more consistent datasets) because missing values, mathematical expressions. Through the development of the
field measurements errors, and significant gap in observa- GEP model, evolut​ ionar​ y comput​ ing recruits the populations
tions can remarkably affect the performance of AI mod- of individuals, and afterward, individuals are selected on the
els. Moreover, collecting, processing, and quality check- basis of their goodness. In addition, genet​ic opera​tors are
ing observed data can be time-consuming and expensive. needed to be adjusted during the performance of the GEP
Using synthetic data generated by TF model can be a faster model. The relationships given by the GEP model include
and more cost-effective approach to testing and develop- two genes at least. Finally, the operator of “addition” links
ing AI models. The TF model is capable of producing existing genes of the GEP model.
hydrological datasets by computation of various statistical This research recruits GeneXproTools software in order to
measures such as the average value, standard deviation present the 100 performances of the GEP model for stream-
(STD), correlation coefficient (CC), coefficient of regres- flow forecast. Through the training and testing stages, the
sion (RC), and coefficient of variance. The basic math- GEP models performance was evaluated by the fitness func-
ematical expression of TF model is obtained to produce tion, namely Mean Square Error (MSE). The results of the
streamflow dataseries as, best GEP performances were obtained among 100 runs for
( ) ( )0.5 the streamflow forecast. For the best performance of GEP
SF i+1 = SF j+1 + RCj SF i − SF j + ti STDj+1 1 − CC2j model, MSE value obtained in the training stage was 6.21,
and additionally, four basic mathematical operators were
(1)
{+,−,÷,×}.
in which SF and SF denote the streamflow and its aver- The setting parameters of GEP performance were
age, respectively. Subscripts of j+1 and i+1 indicate the �
exp(x), ln(x), 1x , x2 , x, min (x, y), max (x, y), avg(x, y), Arctan(x), tanh (x), 1− x ,num-
3
√ � ��

number of month. In addition, RCj is computed by SF val-

ber of generation=999, maximum number of genes=3,
ues at month j+1 and month j. Superscript of i used in SF
mutation rate=0.00138, number of chromosomes= 30, and
expresses the month in which synthetic data generation get
MSE=1486.65.
starts and ti is the number i related to a random number in
Among the 100 runs of GEP model, the best performance
a SF dataseries following trend of the normal distribution.
is expressed
CCj is computed for SF on months j and j+1. Moreover, j
denotes the number of month from January to December.
�� �2 � �2 �
(Q(t − 3) + Q(t − 2)) Q(t − 2)
Q(t) = Q(t − 1) + ×
Finally, i denotes j, month 12+ j, ..., month n + j. Q(t − 3) Q(t − 1)

Once the 100 replicates of the predictors were produced +

√
3
Q(t − 1) × {min (Q(t − 1), Q(t − 3)) − Q(t − 2) − [Q(t − 2) − Q(t − 1)]}
as the training patterns, the training and testing stages of (2)
AI models have been performed. The 95% confidence
intervals have been ultimately defined with aid of obtain-
ing the 2.5th and 97.5th percentiles of the generated dis- Multivariate Adaptive Regression Spline (MARS)
tribution. In this study, the statistical model of forward
selection, introduced as stepwise regression, was utilized MARS model provides interpretable expressions that are devel-
to select the optimum number of lags time that is capable oped by using cross-validation, splines, and non-parametric
of plummeting complexity of AI models. Ten various com- analysis (Friedman 1991). In the case of hydrological applica-
binations [Q(t−1) to Q(t−10)] were examined by forward tions, previous investigations have discovered that the MARS
selection. In each step of forward selection methodology, model had the best performance when splines used in the MARS
the one variable is added to give the single best improve- structure formed the second-order polynomial for analyzing
ment to desired regression model. The first-three lags time various problems such streamflow predictions (e.g., Al-Sudani
[Q(t−1), Q(t−2), Q(t−3)] have been selected as the best et al. 2019; Adnan et al. 2020; Amininia and Saghebian 2021)
combination of input variables for MT, MARS, and EPR and daily forecast of dissolved oxygen concentration (Heddam
models. 2018). The complexity of the regression equation given by the
MARS model is generally evaluated during the forward and
Gene‑Expression Programming (GEP) backward steps. As the forward pass stage is run, a polynomial
regression equation, composed of the sets of bias term and basis
GEP, introduced as a new branch of Genetic Algorithm functions (BFs), is developed. Equation (3) denotes the general
(GA), is capable of producing mathematical expressions structure of the MARS model for streamflow forecast:
in order to provide a reasonable interpretation of dataset
patterns (Ferreira 2001; Ferreira 2006). GEP is composed NBF
∑
of many sub-trees that are generated from well-trained Q(t) = 𝜓0 + 𝜓j .BF j {Q(t − 1), Q(t − 2), Q(t − 3)} (3)
j=1

13
84482 Environmental Science and Pollution Research (2023) 30:84474–84490

in which ψ0, ψj, BF, and NBF are the bias, the constant coef- This study recruits the pruned and smoothed structures
ficients related to basis functions, the basis functions, and the of M5MT. The most efficient result of M5MT gave the most
number of basis functions, respectively. preferable MSE values in training (MSE=1353.43) and test-
The performance of MARS model is appraised by the Gen- ing (MSE=1958.33) phases in the estimation of Q(t) among
eralized Cross-Validation (GCV) criterion that prevents the 100 runs of M5MT. In this way, the multilinear equation
possibility of overfitting. The MARS model was implemented along with one rule was obtained as:
in MATLAB software. In the case of the best performance
Q(t) = 31.0332 − (0.0966 × Q(t − 3))
of MARS model, the GCV value and total effective number
of parameters were obtained 1462 and 8.5, respectively. This − (0.2613 × Q(t − 2)) + (1.0781 × Q(t − 1)) (6)
study used Particle Swarm Optimization (PSO) algorithm to
yield the weighting coefficients of MARS model. Through
100 performances of MARS model, Eq. (4) results in the most Performance assessment of soft computing
accurate forecast of streamflow: models

Q(t) = 55.899 + 1.708 × Max (0, Q[t − 3] − 69.27) + 0.1841

× Max (0,151.1 − Q[t − 1]) − 0.0015644 (4)
× Max (0, Q[t − 3] − 98.654) × Max (0, Q[t − 1] − 90.538)

Performances of soft computing models

In Eq. (4), the MSE value given by training of PSO algo-
rithm is 0.636. This section evaluates the performance of soft computing
models, i.e., GEP, MARS, and M5MT, to predict river flow
M5 Model Tree for training and testing stages of historical data. The best
performance of each model is selected based on three well-
M5 Model Tree, initially implemented by Quinlan (1992), known statistical indices (i.e., coefficient of correlation [R],
is considered a generalized form of regression decision root mean square error [RMSE], and mean absolute error
trees. Each tree consists of nodes and leaves in a way that [MAE]). To make a robust judgment about the performance
multivariate linear regression technique is performed in of models, overfitting of models performance is considered.
the leaf (terminal) nodes of the tree. The M5 version of This means that results of the training phase should not have
MT utilizes a divide-and-conquer technique in order to become too much accurate whereas the performance of test-
solve the problems by splitting the space search of the ing phase stands for remarkably low accuracy.
problem into many sub-spaces (sub-problems). In MT, As demonstrated in Table 2, the statistical indices for
Standard Deviation Reduction (SDR) of the target values MARS model provide flow forecasts with a more slightly
at each node is employed to evaluate a splitting criterion accurate (RMSE=36.765m 3/s and MAE=24.259 ­m 3/s)
for the structure tree (Solomatine and Xue 2004; Etemad- in the training stage than M5MT (RMSE=36.789 ­m 3/s
Shahidi and Bali 2012; Shahabi and Azarpira 2021). To and MAE=24.381 ­m3/s) and GEP (RMSE=38.556 ­m3/s
reach the best performance of MT, there are two operators, and MAE=25.132 ­m3/s) models. R values of the training
namely pruning and smoothing processes, which can be performance indicate that MARS (R=0.841) and M5MT
traded off during training the MT. (R=0.841) models have comparatively the same efficiency
This research uses Weka3.9 software to implement in the monthly flow forecast. Additionally, GEP model
M5MT for the evaluation of streamflow forecast. M5MT is (Eq. (2)) stands for relatively lower accuracy level with R
implemented for three-lagged streamflow variables to pro- of 0.839. In the testing stage, all statistical indices indi-
vide multilinear regression models. For instance, M5MT cate that M5MT with multivariate linear equation provides
presents the following relationship to predict Q(t): the best performance (R=0.870, RMSE=44.253 ­m3/s, and
MAE=28.424 ­m3/s) when compared with GEP (R=0.865,
Q(t) = a0 + a1 × Q(t − 1) + a2 × Q(t − 2) + a3 × Q(t − 3)
RMSE=46.243 ­m3/s, and MAE=28.940 ­m3/s) and MARS
(5)
(R=0.848, RMSE=47.498 ­m3/s, and MAE=30.998 ­m3/s).
in which a 0 is the bias term and a 1 to a 3 are weight- As seen in Table 2, in the training phase, MARS and
ing coefficients that are computed by the least square M5MT performed almost equally well while in the testing
technique.

13
Environmental Science and Pollution Research (2023) 30:84474–84490 84483

Table 2  Statistical indices during the training and testing stages of To deeply explain the performance of AI models, the
the AI models for historical period analysis of the violin plots was utilized in Fig. 6, depicting
Model Phases R RMSE ­[m3/s] MAE ­[m3/s] the distribution of Relative Error (RE) values for all predic-
tive models, namely MT, MARS, and GEP. From Fig. 6, it
GEP Training 0.839 38.556 25.132
can be inferred that the violin plot by GEP demonstrated a
Testing 0.865 46.243 28.94
narrower range of RE values and additionally the GEP plot
MARS Training 0.841 36.765 24.259
is wider close to zero in comparison with plots of MT and
Testing 0.848 47.493 30.998
MARS models. On the other hand, GEP violin plot indicates
M5MT Training 0.841 36.789 24.381
that a greater number of data points have RE values close to
Testing 0.870 44.253 28.424
zero compared with MARS and MT models. RE distribution
in the three AI models was rather the same. Additionally,
phase, M5MT performed slightly better than GEP. Fig- the median and interquartile range (IQR) values of the RE
ure 5 also illustrates the performance of GEP, MARS, and are obtained relatively the same for the three AI models.
M5MT models during training and testing stages. For approximately RE=0.1, the maximum width of GEP

Fig. 5  Forecasted and observed

streamflow during the train-
ing and testing stages for the
(a) GEP, (b) MARS, and (c)
M5MT models

13
84484 Environmental Science and Pollution Research (2023) 30:84474–84490

Fig. 6  Violin plots of RE values

for all predictive models

violin graph had higher frequency than those obtained by Compared with the results of training and testing performances,
maximum frequencies of MARS and MT models. quite a few performances of MARS (10 out of 100) suffered
from overfitting owing to high differences between R values
Stochastic data generation and uncertainty analysis of training and testing phases. For instance, the results of 93rd
run of MARS model yielded R values of 0.9 and 0.01 for the
In order to perform the uncertainty of model inputs dur- training and testing stages, respectively. On the contrary, only
ing the streamflow forecast, MCS approach involving the 58th performance of GEP model indicated R values of 0.19 and
stochastic generation of hundred streamflow sets was used. 0.83 in the training and testing stages, respectively.
In this regard, AI models (i.e., GEP, MARS, and M5MT) A significant trade-off between the training and test-
have been implemented in a MCS-based framework. The ing results demonstrated that M5MT model predicted
MSC technique is carried out by the repeated generation flow discharge with slightly higher accuracy in compar-
of random parameters from their desired probability dis- ison with GEP model. In Fig. 7, the R values given by
tributions, and then, the statistics of the output variable the M5MT and GEP models in the testing phases ranged
[Q(t)] are obtained. In addition, with aid of TF methodol- from 0.62 to 0.87 (0.62 to 0.87 for training stage) and
ogy performance, the input variables [Q(t−1), Q(t−2), and 0.60 to 0.85 (0.62 to 0.87 for training stage), respectively.
Q(t−3)] were randomly resampled without replacement Due to a large number of overfitting occurrences, the
for 100 repetitions, preserving the statistical properties MARS model stood at the lowest level of accuracy in the
of historical and generated datasets. The 95% confidence monthly flow forecast, resulting in ranges of R values
interval of estimation was then determined since this con- 0.68–0.92 and −0.2 to 0.82 for the training and testing
fidence interval provides more information compared with stages, respectively.
other statistical information about the range of predictions
related to the AI models. The 95% confidence intervals Uncertainty analysis of soft computing models
are determined by finding the 2.5th and 97.5th percen-
tiles of the constructed distribution (Anvari et al. 2022). To assess the performance of the GEP, MT, and MARS
Through all distribution functions, results indicated that models under the uncertainty conditions and also to quan-
both Normal and Log-normal distributions were the best- tify the uncertainty, two indices of P-factor and R-factor
fitted functions for the streamflow datasets. Figure 7 illus- were used. The P-factor demonstrates the percentage of
trates training and testing performances of the GEP and tested data that is bracketed by the 95% forecast boundary
M5MT models through 100 random runs for streamflow and referred to as 95PPU prediction uncertainty (95PPU).
prediction at time t. The 95PPU is yielded based on the 2.5% and 97.5% lev-
As seen in Fig. 7, the quantity of runs with R value≥0.80 for els of the cumulative distribution of the variables which
flow forecast during testing stages was 50, 53, and 23 out of 100 are outputted from GEP, MT, and MARS models, here
data patterns for GEP, M5MT, and MARS models, respectively. streamflow forecasts. The goodness-of-fit of uncertainty

13
Environmental Science and Pollution Research (2023) 30:84474–84490 84485

Fig. 7  Illustration of AI models

performances during train-
ing and testing of 100 random
runs for streamflow prediction
at time t, Q(t): (a) GEP, (b)
M5MT, and (c) MARS

13
84486 Environmental Science and Pollution Research (2023) 30:84474–84490

measures is calculated based on the distance between the judge the performance of the best model under conditions
lower and upper limits of 95PPU band and the percentage of uncertainty, both the minimum R-factor index and the
of outputted data located inside the 95PPU band. The aver- maximum 95%PPU should be taken into account at the
age distance between the lower and the upper 95PPU band same time, and according to the mentioned cases, it can
is characterized by d : be said that the performance of the M5MT model under
conditions of uncertainty is better than other models. The
1 ∑n (
(7) other AI models have been better.
)
d= Xu − XL t
n t=1
Reducing the uncertainty of AI models requires a com-
where n is the total measured data and XL and Xu cor- bination of strategies that involve improving the observed
respond to the upper and the lower limits of 95PPU band, data, algorithm, interpretability potential of AI models,
i.e., 97.5% and 2.5% levels, respectively. When the 100% of and validation of the AI model, as well as incorporating
measured data is bracketed by the 95PPU, d becomes close human feedback when necessary. In this case, uncertainty
to zero and the best result is achieved. of flow river forecast can come from disparate sources
R-factor, as expressed below, is considered another such as various combinations of input variables, setting
index to quantify the robustness of the uncertainty analy- parameters of technical model uncertainty, structural
sis. This measure varies in the range of [0, 1] and is deter- function of AI models, and typical field measurements
mined as the average band width of the 95PPU divided by of flow river. Applying various fields studies to meas-
the standard deviation (STD) of the measured streamflow ure/observe flow river is generally out-of-hand of schol-
data points. A desirable value for the R-factor [ d∕STD ] is ars, and consequently, its uncertainty can be discarded.
less than 1 (e.g., Abbaspour et al. 2007; Yang et al. 2008; Additionally, the selection of appropriate lags of time
Singh et al. 2014). series flow rivers plays an important role in increasing
This investigation uses P-factor and R-factor to evalu- (or decreasing) accuracy level of AI models (i.e., GEP,
ate the uncertainty analysis of the AI models forecast. M5MT, and MARS). The sufficient number of lags time
More specifically, the value of the P-factor varies from was computed by mathematical models, namely the Auto-
0 to 1 whereas the valid range of R-factor is between 0 Correlation Function (ACF) and the Partial ACF (PACF),
and infinity. Once P-factor and R-factor were obtained 1 as can be seen in Saghafian et al. (2013). By the way,
and zero, respectively, the best performance of AI models using number of lags time can plummet uncertainty of
is met in terms of uncertainty quantification. From the AI models performance. Moreover, the setting param-
two mentioned factors, the results of AI models uncer- eters of AI models affect uncertainty quantification. In
tainty always seek to bracket the majority of the observed M5MT, the usability of the smoothed and pruned modes
dataset, obtaining large value of P-factor and the lowest makes changes in the number of if-then rules, simplifying
value of R-factor. Figure 8 depicts that 95% confidence multivariate expression given by M5MT. This research
interval of the predicted values of Q(t) has been defined employed these modes to stand the driven equations at
owing to the fact that confidence interval obtains more their utmost level of performance. In GEP model, one of
useful information in comparison with other statistical the leading factors affecting the performance of uncer-
measures/tests in relation with the range of uncertainty tainty quantification is by selecting typical optimization
analysis. As depicted in Fig. 8, once the confidence strategies in producing the genes with various genetic
interval gets wider, the precision level of the forecasted operators. The present research applied optimal evolution,
streamflow will get lower. reaching the most accurate flow forecast for all 100 runs
Uncertainty results of MARS, GEP, and MT models (patterns of datasets). Furthermore, interaction degree of
have been calculated and quantified for predicting the MARS expression cannot be considered as one of uncer-
flow of the Karoon River 1 month later according to tainty sources if interaction degree of input variables (i.e.,
Table 3. Smaller values of R-factor, which represents the Q(t−1), Q(t−2), Q(t−3)) is kept constant. In this inves-
uncertainty bandwidth, indicate good performance of the tigation, the maximum degree of MARS-obtained poly-
AI model. From this point of view, the M5MT model has nomial is two that takes advantages such as simplifying
the smallest R-factor value compared to other models. expression and reaching accurate predictions. As a merit,
On the other hand, the 95% PPU values that represent applying GCV parameter can substantially reduce over-
the percentage of definitive data (historical observations) fitting occurrence (or complexity of adaptive regression
are in the 95% uncertainty estimation band; for the men- provided in the forward phase of MARS development),
tioned models, it is between 59.5 and 64%. Finally, to plummeting uncertainty level in the flow river forecast.

13
Environmental Science and Pollution Research (2023) 30:84474–84490 84487

Fig. 8  The 95% confidence

interval of Q(t) predictions in
(a) GEP, (b) MARS, and (c)
M5MT models for uncertainty
analysis

Conclusion
Table 3  Comparison of the Models 95% PPU R-factor The present research employed 336 monthly streamflows in
uncertainty results of soft
computing models in predicting Armand station located at Karoon watershade, southwest of
GEP 0.64 2.03
the Karoon River streamflow Iran, in order to determine the uncertainty levels related to
MARS 0.61 1.92
M5MT, GEP, and MARS models. AI models, at first, were
M5MT 0.59 1.67
run 100 times with aid of data patterns generated by MC

13
84488 Environmental Science and Pollution Research (2023) 30:84474–84490

methodology. Next, the best performance of each AI model Data Availability Some or all data, models, or codes that support the
was statistically evaluated. The MARS and M5MT mod- findings of this study are available from the corresponding author upon
reasonable request.
els showed better accuracy in flow forecast results than the
GEP models during the training phase, as demonstrated by Declarations
statistical measures such as R, RMSE, and MAE. Moreover,
during the testing phase, the M5MT model’s multivariate Ethical approval All procedures performed in studies involving
linear regression equations performed better than those of human participants were in accordance with the ethical standards
of the institutional and/or national research committee and with the
the GEP and MARS models. Moreover, the study examined 1964 Helsinki declaration and its later amendments or comparable
the uncertainty of GEP, MT, and MARS in predicting the ethical standards.
monthly streamflow. According to 95%PPU and R-factor, the
results indicated that MT had the least uncertainty compared Consent to participate Informed consent was obtained from all indi-
vidual participants included in the study.
to MARS and GEP models.
Overall, the study demonstrated significance of uncer- Consent for publication All the authors give the publisher the permis-
tainty related to these artificial intelligence models com- sion of the authors to publish the research work.
monly employed for prediction of flow discharge. Since
Competing interests The authors declare no competing interests.
evaluation of river flow, particularly in extensive drainage
basins, plays a substantial role in management of water
demands for various purposes, uncertainty in its fore-
cast leads to unrealistic values of flow discharge being
References
assessed by the AI models during drought and flood
events. In this way, the present findings obviously attach Al-Sudani ZA, Salih SQ, Sharafati A, Yaseen ZM (2019) Development
the importance to the uncertainty analysis of flow dis- of multivariate adaptive regression spline integrated with differen-
charge forecast models as an imperative case that had tial evolution model for streamflow simulation. J Hydrol 573:1–12
Abbaspour KC (2008) SWAT-CUP user manual. Federal Institute of
not already been cornerstone of hydrologic investiga-
Aquatic Science and Technology (Eawag), Swiss
tions. Thus, it is suggested that the present methodol- Abbaspour KC, Yang J, Maximov I, Siber R, Bogner K, Mieleitner
ogy assessing the uncertainty level of these AI models J, Srinivasan R (2007) Modelling hydrology and water quality
is considered a primary case of traditional investigations in the pre-alpine/alpine Thur watershed using SWAT. J Hydrol
333(2-4):413–430
(e.g., hydraulic modeling of floods in natural streams and
Adnan RM, Liang Z, Heddam S, Zounemat-Kermani M, Kisi O, Li
flood routing in watershed basins by hydrological model) B (2020) Least square support vector machine and multivariate
on the river flow forecast. Due to the significant role of adaptive regression splines for streamflow prediction in mountain-
the streamflow in various aspects of water resource man- ous basin using hydro-meteorological data as inputs. J Hydrol
586:124371
agement, underestimation of the daily streamflow val-
Adeloye AJ (2012) Hydrological sizing of water supply reservoir. In:
ues could ecologically put the quality of surface water Bengtsson L, Herschy RW, Fairbridge RW (eds) Encyclopedia of
resources in danger. An accurate flow forecast in rivers lakes and reservoirs. Springer, Dordrecht, pp 346–355
would permit the policy-makers consider well-suited per- Afkhami M, Shariat M, Jaafarzadeh N, Ghadiri H, Nabizadeh
R (2007) Developing a water quality management model
spectives for ameliorating a broad range of issues such as
for Karoon and Dez Rivers. J Environ Health Sci Eng
rivers management, reliable design of water-conveying 4(2):99–106
structures, flood monitoring, and good management of Amininia K, Saghebian SM (2021) Uncertainty analysis of monthly
water allocation for agricultural affairs. river flow modeling in consecutive hydrometric stations using
integrated data-driven models. J Hydroinf 23(4):897–913
Overall, the proposed methodologies proved that pieces
Anvari S, Rashedi E, Lotfi S (2022) A Coupled Metaheuristic Algo-
of important information can be useful for good manage- rithm and Artificial Intelligence for LongLead Stream Flow Fore-
ment of the streamflow and possibilities for forthcoming casting. International Journal of Optimization in Civil Engineer-
improvements. ing 12(1):91–104
Anvari S, Mousavi SJ, Morid S (2014) Sampling/stochastic dynamic
programming for optimal operation of multi-purpose reservoirs
using artificial neural network-based ensemble streamflow predic-
Author contribution Mohammad Najafzadeh: providing computer
tions. J Hydroinf 16(4):907–921
programming codes for AI models, performing Artificial Intelligence
Aqil M, Kita I, Yano A, Nishiyama S (2007) A comparative study of
models for stream flow forecast, writing analysis of AI models results,
artificial neural networks and neuro-fuzzy in continuous modeling
improving introduction, and writing literature review, abstract, conclu-
of the daily and hourly behaviour of runoff. J Hydrol 337:22–34
sion, and comparisons sections. Sedigheh Anvari: writing introduc-
Bayazit M (2015) Nonstationarity of hydrological records and recent
tion, data analysis, case study, describing uncertainty descriptions, and
trends in trend analysis: a state-of-the-art review. Environ Process
performing uncertainty analysis of AI models, and providing figures
2(3):527–542
and tables.

13
Environmental Science and Pollution Research (2023) 30:84474–84490 84489

Bazartseren B, Hildebrandt G, Holz KP (2003) Short-term water level Marcé R, Comerma M, García JC, Armengol J (2004) A neuro-fuzzy
prediction using neural networks and neuro-fuzzy approach. Neu- modeling tool to estimate fluvial nutrient loads in watersheds
rocomputing 55:439–450 (can be deleted) under time-varying human impact. Limnol Oceanogr Methods
Bensoussan A, Farhi N (2010) Uncertainties and Risks in Water 2(11):342–355
Resources Management. In: The economics of sustainable devel- Modaresi F, Araghinejad S, Ebrahimi K (2018) A comparative assess-
opment (Chapter: Uncertainties and Risks in Water Resources ment of artificial neural network, generalized regression neural
Management). Publisher: Economica network. Least-Square support vector regression, and K-nearest
Cheng CT, Feng ZK, Niu WJ, Liao SL (2015) Heuristic methods for neighbor regression for monthly Streamflow forecasting in linear
reservoir monthly inflow forecasting: A case study of Xinfengji- and nonlinear conditions. Water Resour Manag 32(1):243–258
ang Reservoir in Pearl River, China. Water 7(8):4477–4495 Montanari A, Grossi G (2008) Estimating the uncertainty of hydro-
Coulibaly P, Anctil F, Bobée B (2000) Daily reservoir inflow fore- logical forecasts: A statistical approach. Water Resour Res
casting using artificial neural networks with stopped training 44(12):W00B08
approach. J Hydrol 230:244–257 Moghaddasi M, Anvari S, Akhondi N (2022) A trade-off analysis of
Cui Q, Wang X, Li C, Cai Y, Liang P (2016) Improved Thomas- adaptive and non-adaptive future optimized rule curves based on
Fiering and Wavelet Neural Network Models for Cumulative Error simulation algorithm and hedging rules. Theor Appl Climatol
Reduction in Reservoir Inflow Forecast. J HydroEnvironment Res 148(1–2):65–78
13:134–143 Narsimlu B, Gosain AK, Chahar BR, Singh SK, Srivastava PK (2015)
Diomede T, Marsigli C, Nerozzi F, Paccagnella T, Montani A (2006) SWAT model calibration and uncertainty analysis for streamflow
Quantifying the discharge forecast uncertainty by different prediction in the Kunwari River Basin, India, using sequential
approaches to probabilistic quantitative precipitation forecast. uncertainty fitting. Environ Process 2(1):79–95
Adv Geosci 7:189–191 Nayak PC, Sudheer KP, Rangan DM, Ramasastri KS (2004) A neuro-
Dehghani M, Saghafian B, Nasiri Saleh F, Farokhnia A, Noori R (2014) fuzzy computing technique for modeling hydrological time series.
Uncertainty analysis of streamflow drought forecast using artifi- J Hydrol 291:52–66
cial neural networks and Monte-Carlo simulation. Int J Climatol Niu WJ, Feng ZK, Cheng CT, Zhou JZ (2018) Forecasting daily
34(4):1169–1180 runoff by extreme learning machine based on quantum-behaved
Di Baldassarre G, Montanari A (2009) Uncertainty in river discharge particle swarm optimization. J Hydrol Eng 23(3):04018002
observations: a quantitative analysis. Hydrol Earth Syst Sci Dis- Noorbeh P, Roozbahani A, Moghaddam HK (2020) Annual and
cuss 6:39–61 monthly dam inflow prediction using Bayesian networks. Water
El-Shafie A, Noureldin A (2011) Generalized versus non-generalized Resour Manag 34(9):2933–2951
neural network model for multi-lead inflow forecasting at Aswan Noori R, Karbassi AR, Moghaddamnia A, Han D, Zokaei-Ashtiani
High Dam. Hydrol Earth Syst Sci 15(3):841–858 MH, Farokhnia A, Gousheh MG (2011) Assessment of input vari-
Etemad-Shahidi A, Bali M (2012) Stability of rubble-mound break- ables determination on the SVM model performance using PCA,
water using H50 wave height parameter. Coast Eng 59(1):38–47 Gamma test, and forward selection techniques for monthly stream
Ferreira C (2001) Gene expression programming: A new adaptive algo- flow prediction. J Hydrol 401(3-4):177–189
rithm for solving problems. Complex Systems 13:87–129 Noori R, Khakpour A, Omidvar B, Farokhnia A (2010) Comparison of ANN
Ferreira C (2006) Gene Expression Programming, 2nd edn. Springer-Verlag, and principal component analysis-multivariate linear regression mod-
Berlin Heidelberg, The Netherlands els for predicting the river flow based on developed discrepancy ratio
Friedman JH (1991) Multivariate adaptive regression splines. Ann Stat statistic. Expert Syst Appl 37(8):5856–5862
19:1–67 Pham T (2011) Tracking the uncertainty in streamflow prediction
He Z, Wen X, Liu H, Du J (2014) A comparative study of artificial neu- through a hydrological forecasting system, [Link] Thesis. Univer-
ral network, adaptive neuro fuzzy inference system and support sity of Twente
vector machine for forecasting river flow in the semiarid mountain Pablo AM, McPhee J, Vargas X (2012) Uncertainty in flood forecasting:
region. J Hydrol 509:379–386 A distributed modeling approach in a sparse data catchment. Water
Heddam S, Kisi O (2018) Modelling daily dissolved oxygen concentra- Resour Res 48(9):W09532
tion using least square support vector machine, multivariate adap- Poul AK, Shourian M, Ebrahimi H (2019) A comparative study of
tive regression splines and M5 model tree. J Hydrol 559:499–509 MLR, KNN, ANN and ANFIS models with wavelet trans-
Hussain D, Khan AA (2020) Machine learning techniques for monthly form in monthly stream flow prediction. Water Resour Manag
river flow forecasting of Hunza River, Pakistan. Earth Sci Inform 33(8):2907–2923
13:939–949 Quinlan JR (1992) Learning with continuous classes. In: Proceedings
Kisi O (2005) Daily river flow forecasting using artificial neural of the 5th Australian Joint Conference on Artificial Intelligence
networks and auto-regressive models. Turk J Eng Environ Sci (AI ’92). World Scientific, Singapore, pp 343–348
29:9–20 Riahi-Madvar H, Dehghani M, Memarzadeh R, Gharabaghi B (2021)
Kroese DP, Brereton T, Taimre T, Botev ZI (2014) Why the Monte Short to long-term forecasting of river flows by heuristic optimi-
Carlo method is so important today. Wiley Interdiscip Rev Com- zation algorithms hybridized with ANFIS. Water Resour Manag
put Stat 6(6):386–392 35(4):1149–1166
Leandro J, Bander A, Beg MNA, Bhola P, Konnerth I, Willems W, Rubinstein RY, Kroese DP (2016) Simulation and the Monte Carlo
Carvalho R, Disse M (2019) Forecasting upper and lower uncer- method. John Wiley & Sons
tainty bands of river flood discharges with high predictive skill. Singh A, Imtiyaz M, Isaac RK, Denis DM (2014) Assessing the perfor-
J Hydrol 576:749–763 mance and uncertainty analysis of the SWAT and RBNN models
McMahon TA, Adeloye AJ (2005) Water Resources Yield. Water for simulation of sediment yield in the Nagwa watershed, India.
Resources Publications, Littleton, CO, USA Hydrol Sci J 59(2):351–364
Moosavi V, Gheissori Fard Z, Vafakhah M (2022) Which one is more Shahabi S, Azarpira F (2021) Evaluating the capability of hybrid data-
important in daily runoff forecasting using data driven models: driven approaches to forecast monthly streamflow using hydro-
Input data, model type, preprocessing or data length? J Hydrol metric and meteorological variables. J Hydroinf 23(6):1165–1181
606:127429

13
84490 Environmental Science and Pollution Research (2023) 30:84474–84490

Steenbergen NV, Willems P (2015) Uncertainty decomposition and Yu P-S, Tseng T-Y (1996) A model to forecast flow with uncertainty
reduction in river flood forecasting: Belgian case study. Flood analysis. Hydrol Sci J 41(3):327–344
Risk Manag 8(3):263–275 Yang T, Asanjan AA, Welles E, Gao X, Sorooshian S, Liu X (2017)
Saghafian B, Anvari S, Morid S (2013) Effect of Southern Oscilla- Developing reservoir monthly inflow forecasts using artificial
tion Index and spatially distributed climate data on improving intelligence and climate phenomenon information. Water Resour
the accuracy of Artificial Neural Network, Adaptive Neuro-Fuzzy Res 53(4):2786–2812
Inference System and K-Nearest Neighbour streamflow forecast- Yang J, Reichert P, Abbaspour KC, Xia J, Yang H (2008) Comparing
ing models. Expert Syst 30(4):367–380 uncertainty analysis techniques for a SWAT application to the
Schnier S, Cai XM (2014) Prediction of regional streamflow frequency Chaohe Basin in China. J Hydrol 358:1–23
using model tree ensembles. J Hydrol 517:298–309 Yaseen ZM, Naganna SR, Sa’adi, Z, Samui, P., Ghorbani, M.A., Salih,
Setegn SG, Srinivasan R, Melesse AM, Dargahi B (2010) SWAT model S.Q., Shahid, S. (2020) Hourly river flow forecasting: application
application and prediction uncertainty analysis in the Lake Tana of emotional neural network versus multiple machine learning
Basin, Ethiopia. Hydrol Process 24(3):357–367 paradigms. Water Resour Manag 34:1075–1091
Shiri J (2018) Improving the performance of the mass transfer-based Yaseen ZM, El-Shafie A, O., Jaafar, H. Afan, A., Sayl. M. N. (2015)
reference evapotranspiration estimation approaches through a cou- Artificial intelligence based models for stream-flow forecasting:
pled wavelet-random forest methodology. J Hydrol 561:737–750 2000-2015. J Hydrol 530:829–844
Soundharajan BS, Adeloye AJ, Remesan R (2016) Evaluating the vari- Yan J, Lio G-Y, Gebremichael M, Shedd R, Vallee DR (2012) Char-
ability in surface water reservoir planning characteristics during acterizing the uncertainty in river stage forecasts conditional on
climate change impacts assessment. J Hydrol 538:625–639 point forecast values. Water Resour Res 48(12):W12509
Solomatine DP, Xue Y (2004) M5 Model Trees and Neural Networks: Zhang X, Liang F, Srinivasan R, Van Liew M (2009) Estimating uncer-
Application to Flood Forecasting in the Upper Reach of the Huai tainty of streamflow simulation using Bayesian neural networks.
River in China. J Hydrol Eng 9(6):491–501 Water Resour Res 45(2):W02403
Talebizadeh M, Morid S, Ayyoubzadeh SA, Ghasemzadeh M (2010) Zealand CM, Burn DH, Simonovic SP (1999) Short term stream-
Uncertainty analysis in sediment load modeling using ANN and flow forecasting using artificial neural networks. J Hydrol
SWAT model. Water Resour Manag 24(9):1747–1761 214(1-4):32–48
Tikhamarine Y, Souag-Gamane D, Ahmed AN, Kisi O, El-Shafie
A (2020) Improving artificial intelligence models accuracy for Publisher’s note Springer Nature remains neutral with regard to
monthly streamflow forecasting using grey Wolf optimization jurisdictional claims in published maps and institutional affiliations.
(GWO) algorithm. J Hydrol 582:124435
Wanger T, Gupta H, Yatheendradas S, Goodrich D, Unkrich C, Schaf- Springer Nature or its licensor (e.g. a society or other partner) holds
fner M (2007) Understanding sources of uncertainty in flashflood exclusive rights to this article under a publishing agreement with the
forecasting for semi-arid regions. Quantification and Reduction author(s) or other rightsholder(s); author self-archiving of the accepted
of Predictive Uncertainty for Sustainable Water Resources Man- manuscript version of this article is solely governed by the terms of
agement. In: Proceedings of Symposium HS2004 at IUGG2007, such publishing agreement and applicable law.
Perugia. IAHS Publication

13