Daily Reservoir Evaporation Estimation Using MLP and ANFIS: A Comparative Study for Sustainable Water Management

Abstract

Reservoir evaporation is a vital component of the hydrological cycle and presents considerable challenges for sustainable water management, especially in arid and semi-arid regions. This study assesses the effectiveness of two Artificial Intelligence (AI) methods: Multilayer Perceptron (MLP) and Adaptive Neuro-Fuzzy Inference System (ANFIS), a combination ANN with fuzzy logic, in estimating daily evaporation from a large reservoir in a semi-arid region. Using eight years of hydrometeorological data from a nearby station, the study employed the ReliefF algorithm as a feature selection method for relevant input variables. The dataset was divided into training, validation, and testing subsets with 5% and 10% validation ratios, using four train–test splits of 70:30, 75:25, 80:20, and 85:15. Various training algorithms (e.g., Levenberg–Marquardt) and membership functions (e.g., generalized bell-shaped functions) were tested for both models. MLP consistently outperformed ANFIS on the test sets, showing higher R² and lower RMSE values. In the best-performing 70:30 split, MLP achieved an R² of 0.8069 and RMSE of 0.0923, compared to ANFIS with an R² of 0.3192 and RMSE of 0.2254. The findings highlight the AI-based approaches’ potential to support improved evaporation forecasting and integration into decision support tools for water resource planning amid changing climatic conditions.

1. Introduction

In recent decades, growing concerns over water scarcity and climate variability have intensified the need for precise estimation of hydrological processes, with evaporation emerging as one of the most challenging yet crucial components to model. However, accurately estimating evaporation poses significant challenges due to the complex and nonlinear relationships among meteorological variables [1,2,3]. Traditional empirical formulas often fall short in capturing these intricate interactions, resulting in limited estimation reliability. These shortcomings have led to the extensive application of Artificial Intelligence (AI)-based methods in evaporation modeling processes in recent years.

Recent studies highlight the value of Artificial Intelligence (AI) models in evaporation prediction, revealing that these shortcomings have led to the widespread application of AI-based methods in evaporation modeling processes. For instance, Jayasinghe et al. (2022) [4] employed deep learning models such as LSTM for pan evaporation forecasting and found that recurrent architectures significantly improved temporal accuracy over traditional methods. Architectures such as LSTM and GRU consistently outperform traditional models in hydrological time series prediction, including runoff and water level forecasting [5,6,7,8]. In a related study, Sharma et al. (2020) [9] showed that Random Forest effectively predicts pan evaporation and outperforms artificial neural networks over long-term datasets, while Bilali et al. (2022) [10] demonstrated that XGBoost achieves high accuracy and interpretability using tools like SHAP and LIME to identify key climatic variables influencing evaporation. Furthermore, many researchers emphasized the role of AI-integrated remote sensing approaches in predicting evapotranspiration under climate variability scenarios, underscoring the need for models that can generalize across diverse environmental settings [11,12,13]. These advancements confirm that AI offers distinct advantages in modeling complex hydrological processes; however, the relative performance of AI models under different data conditions remains an open question. This motivates a comparative approach to evaluate and optimize AI techniques such as ANNs and ANFIS for operational evaporation forecasting.

Artificial Neural Networks (ANNs), Adaptive Neuro-Fuzzy Inference System (ANFIS), Support Vector Regression (SVR), deep learning models, and various hybrid Machine Learning (ML) techniques are widely employed in this context. Owing to their ability to model nonlinear relationships, these methods demonstrate superior modeling capability over conventional techniques, particularly under uncertain or incomplete data conditions [1,2,3,14].

The literature highlights that ANNs and ANFIS exhibit notably strong modeling potential, especially in scenarios where climatic data are limited [15,16]. Nevertheless, the superiority of one model over another is largely dependent on the characteristics of the dataset, prevailing climate conditions, and the selected input variables. Hybrid and subset-based modeling strategies have proven effective in enhancing ANFIS, in some cases achieving better results than standard models [1,17].

Comparative analyses using ANN and ANFIS models demonstrate that both methods can effectively estimate daily reservoir evaporation [1,2,3,15]. Allawi et al. [18] stated that ANFIS outperformed ANNs in long-term forecasts in evaporation prediction. Accordingly, both are considered powerful modeling tools, and the choice between them may vary depending on the application purpose, data structure, and available computational resources [1,2,3,14,15,16,17,19,20,21]. ANN models are generally distinguished by faster learning processes, better generalization capabilities, and lower computational costs. In contrast, ANFIS models tend to yield enhanced precision, particularly when limited datasets and optimized input variables are used [1,2,20]. In this context, powerful feature selection algorithms like ReliefF significantly enhance the efficiency and accuracy of both ANN and ANFIS models [22,23]. ReliefF identifies the most relevant input features associated with the target variable, eliminating irrelevant and noisy variables and ensuring that the models are trained with only meaningful data. This leads to faster learning and reduces computational demands in ANN models, while in ANFIS models, it reduces the number of fuzzy rules, thereby improving interpretability and generalization ability. Moreover, in cases with small or limited datasets, the rule-based structure of ANFIS becomes much more effective when combined with carefully selected key features [24,25]. Consequently, the proper integration of preprocessing steps such as ReliefF contributes to making these models more reliable and applicable under real-world conditions, enhancing both reliability and operational efficiency [26].

Located in Turkey’s Southeastern Anatolia region, the Atatürk Dam plays a strategic role in regional water management through large-scale irrigation and hydroelectric power generation; however, the use of decision support systems such as machine learning, which could contribute to sustainable water management, remains limited. This study aims to analyze and compare the estimation performance and generalization ability of Artificial Neural Networks (ANNs) in the form of Multilayer Perceptron (MLP) and Adaptive Neuro-Fuzzy Inference System (ANFIS) models for estimating daily evaporation at the Atatürk Dam using long-term hydrometeorological data. Feature selection is performed using the ReliefF algorithm to identify relevant input variables, and the models’ estimation success is evaluated based on statistical metrics under different data partitioning strategies.

2. Materials and Methods

2.1. Study Area and Data Collection

This study was conducted using data from the Atatürk Dam Reservoir, which is situated on the Euphrates River between the provinces of Adıyaman and Şanlıurfa in southeastern Turkey. As one of the central components of the Southeastern Anatolia Project, the dam serves as a key infrastructure for hydroelectric energy production and irrigation. Completed in 1992, the Atatürk Dam has a height of 169 m and forms one of Turkey’s largest reservoirs, with a surface area of approximately 817 km² and a total storage capacity of 48.7 billion cubic meters [27].

Daily hydrometeorological data were obtained from the Turkish State Meteorological Service (TSMS) for the period between 2004 and 2011 to develop predictive evaporation models [28]. The selected meteorological station is near the reservoir and provides consistent, long-term observational records. The dataset spans eight years and includes key variables influencing evaporation such as air temperature (°C), relative humidity (%), wind speed (m/s), sunshine duration (hours), and atmospheric pressure (hPa). Daily pan evaporation measurements obtained from the same station were used as target output data for model training and evaluation.

2.2. Data Preprocessing

Descriptive statistics, including measures such as the mean, median, and standard deviation, were employed to summarize the fundamental characteristics of the dataset and are presented in

Table 1. Descriptive statistics of the dataset.

Parameters		Statistical Data
		Min	Mean	Max	SD	Sample Variance	Kurtosis	Skewness	Units
Input	Min. Air Temperature	−4.30	15.71	30.20	5.95	35.44	−0.62	−0.27	°C
	Mean Air Temperature	4.70	23.95	36.50	6.52	42.55	−0.75	−0.39	°C
	Max. Air Temperature	9.30	31.59	44.60	7.06	49.83	−0.50	−0.52	°C
	Total Precipitation	0.00	0.58	40.60	2.91	8.47	74.49	7.89	mm
	Sunshine Hours	0.00	10.41	14.10	3.24	10.47	1.70	−1.48	hours
	Max Humidity	15.00	61.58	100.00	17.51	306.56	−0.69	0.11	%
	Min Humidity	5.00	27.39	86.00	12.28	150.71	3.42	1.39	%
	Mean Humidity	11.90	42.97	92.30	14.10	198.84	0.31	0.67	%
Output	Total Evaporation	0.00	9.27	19.40	4.29	18.42	−1.02	−0.15	mm

.

The ReliefF algorithm, a statistical feature selection method, was utilized to enhance model performance by identifying and selecting input variables with the strongest relationship to the target variable [29]. The ReliefF feature selection algorithm was applied to rank the importance of input variables and confirm their relevance in evaporation prediction. The selected features were then normalized using min–max scaling to bring all variables into the [0, 1] range (Equation (1)), thus improving the convergence rate of the learning algorithms, reduce inconsistencies, standardize feature ranges, and improve computational stability [30].

$$\mathrm{Y}={\displaystyle \frac{{\mathrm{X}}_{\mathrm{i}}-{\mathrm{X}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}{{\mathrm{X}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{X}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}$$

(1)

Here, X_i represents the raw data, X_min and X_max denote the minimum and maximum values of X, respectively, and Y is the normalized data value. Normalization is a vital preprocessing step in machine learning that transforms data into a consistent scale, making it more suitable for analysis and modeling.

Following feature selection and normalization, a 1712 × 7 matrix was formed, representing the final dataset to be used in model training and evaluation. Each of the 1712 rows corresponds to a sample, and each of the 7 columns to a selected input feature. Prior to model training, all input variables were normalized to ensure uniform scaling, facilitating improved convergence and model stability. The normalized dataset was then divided into training, validation, and testing subsets. Two fixed validation proportions, 5% and 10% of the total dataset, were examined to evaluate the effect of validation size on MLP and ANFIS model performances. This dataset was not used during training but only for validation purposes. For each validation level, the remaining data were partitioned into training and testing sets using various ratios: specifically, training sets comprising 70%, 75%, 80%, and 85% of the nonvalidation data, with the remainder allocated to testing. For ANFIS modeling, the “checking data” option of the MATLAB R2017b (9.3.0.713579) ANFIS toolbox was used to assign validation subsets during training. This feature enabled performance monitoring and validation-based adjustment of the fuzzy inference system. Consequently, both MLP and ANFIS models were trained and validated under comparable conditions using the selected data splits. This flexible partitioning approach enabled a robust comparison of model behavior under different data availability scenarios, supporting a more comprehensive evaluation of model accuracy, generalization ability, and sensitivity to training data volume.

2.3. Model Development

2.3.1. Artificial Neural Network (ANN)

Artificial Neural Networks (ANNs) are widely recognized as robust computational models for knowledge extraction, offering the capacity to infer complex, and often nonlinear, causal relationships between input and output variables. In addition, ANNs enable the identification of hidden neurons and facilitate the analysis of their internal dynamics, thereby enhancing the interpretability of classification processes [31]. As a supervised learning method, ANNs are particularly advantageous due to their adaptability and predictive accuracy in time series forecasting.

In the present study, separate ANN models were constructed for each scenario derived through the ReliefF feature selection algorithm. A feedforward Multilayer Perceptron (MLP) architecture was utilized via MATLAB’s (R2017b (9.3.0.713579)) neural network fitting application. To evaluate model performance, three distinct training algorithms, Bayesian Regularization (BR), Levenberg–Marquardt (L-M), and Scaled Conjugate Gradient (SCG), were applied across various data partitioning schemes as mentioned above. Model performance was iteratively evaluated on the validation set to avoid overfitting. The overall structure of the MLP model used in this study, including input variables, hidden layer configuration, and output layer, is illustrated in Figure 1. Among these, the scenario yielding the most optimal performance determined based on the coefficient of determination (R²) and Root Mean Square Error (RMSE) was selected for final model verification as mentioned in Section 2.4.

2.3.2. Adaptive Neuro-Fuzzy Inference System (ANFIS)

The Adaptive Neuro-Fuzzy Inference System (ANFIS), developed by Jang et al. [32], combines the learning capabilities of Artificial Neural Networks (ANNs) with the reasoning structure of Fuzzy Inference Systems (FISs). This integration enables automatic tuning of fuzzy rules through adaptive networks, improving training efficiency and convergence speed compared to traditional ANNs.

ANFIS modeling proceeds in three main stages. Initially, system parameters, including inputs, outputs, antecedents, and consequents, are defined using the Takagi–Sugeno approach. Fuzzy rules follow the “if–then” structure, where antecedents involve nonlinear membership functions and consequents are expressed as linear functions of inputs or constants. In the second stage, these rules are mapped onto a multilayer adaptive network architecture.

In parallel, an ANFIS model was constructed to incorporate human-like reasoning through fuzzy logic combined with the adaptive learning capacity of neural networks. Separate ANFIS models were constructed for each scenario derived through the ReliefF feature selection algorithm. The scenario demonstrating the highest performance was selected for final model verification, as outlined in Section 2.4.

2.4. Model Evaluation and Comparison

The trained MLP and ANFIS models were evaluated on the test dataset using two standard performance metrics: coefficient of determination (R²) and Root Mean Square Error (RMSE).

RMSE is one of the most widely used indicators for assessing a model’s prediction accuracy. In the formulas below [33], X_i denotes the simulated values, Y_i the observed values, and m the total number of data points.

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{X}}_{\mathrm{i}}-{\mathrm{Y}}_{\mathrm{i}}\right)}^2}$$

(2)

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{m}}{\left({\mathrm{X}}_{\mathrm{i}}-{\mathrm{Y}}_{\mathrm{i}}\right)}^2}{{\sum }_{\mathrm{i}=1}^{\mathrm{m}}{\left(\stackrel{\mathrm{-}}{\mathrm{Y} }-{\mathrm{Y}}_{\mathrm{i}}\right)}^2}}$$

(3)

R-squared, which ranges from 0 to 1, represents the squared correlation coefficient (R²) between the observed and predicted datasets. It indicates the proportion of variance in the dependent variable that is explained by the independent variables, effectively reflecting the model’s goodness of fit. These metrics allowed for both quantitative comparison and error structure analysis.

2.5. Software and Implementation

All model simulations and performance analyses were conducted in MATLAB R2017b (9.3.0.713579). ReliefF and MLP/ANFIS implementations were carried out using MATLAB’s (R2017b (9.3.0.713579)) built-in Neural Network and Fuzzy Logic Toolboxes. Additional validation routines and statistical analysis were implemented using custom scripts.

3. Results

Input Selection and Model Development

As described in Section 2.2, the ReliefF algorithm implemented via MATLAB R2017b (9.3.0.713579) was utilized to identify the most influential meteorological variables affecting evaporation, based on data obtained from the Turkish State Meteorological Service. The output of the algorithm, presented in

Table 2. ReliefF algorithm results.

Parameter Name	ReliefF Ordering	Weights
Sunshine duration (hour)	5	0.0034
Average temperature (°C)	2	0.0016
Total precipitation (mm)	4	0.0014
Maximum temperature (°C)	3	0.0012
Minimum temperature (°C)	1	0.0012
Average humidity (%)	8	0.0005
Min humidity (%)	7	0.0000
Max humidity (%)	6	−0.0002

, includes weight scores assigned to each parameter, where values closer to +1 indicate a stronger positive contribution to the target variable. Variables with the highest weights were selected as inputs for subsequent MLP and ANFIS modeling processes.

Among the evaluated parameters, maximum humidity (%) exhibited a negative weight, signifying an inverse association with evaporation. Furthermore, minimum humidity (%) displayed a similar pattern and was determined to have negligible predictive influence. Based on these findings, the final set of input variables included sunshine duration (hours), average temperature (°C), total precipitation (mm), maximum temperature (°C), minimum temperature (°C), and average humidity (%).

Following the feature selection process using the ReliefF algorithm, all input variables were subjected to normalization to ensure uniform scaling, thereby enhancing model stability and facilitating faster convergence during training. The normalized dataset was subsequently partitioned into training, validation, and testing subsets. To assess the impact of validation set size on model performance, two fixed validation proportions, 5% and 10% of the total dataset, were employed. For each validation level, the remaining data were further split into training and testing sets using varying ratios, wherein 70%, 75%, 80%, and 85% of the nonvalidation data were allocated for training, with the remainder reserved for testing.

This flexible data partitioning strategy enabled a robust assessment of model performance under different data availability conditions, allowing for a comprehensive evaluation of prediction accuracy, generalization capacity, and sensitivity to training data volume. Among the developed MLP and ANFIS models, the optimal configuration selected based on the lowest Root Mean Square Error (RMSE) and highest coefficient of determination (R²) was identified for final model validation and deployment. The detailed model development parameters corresponding to the best-performing MLP and ANFIS configurations are summarized in

Table 3. ANFIS model development parameters.

Model Name	Parameters	Value
MLP	Training Method	L-M
	Number of Epochs	15
	Number of Neurons in Hidden Layer	18
ANFIS	Input Membership Function (MF) Type	gbellmf
	Number of Input MFs	[2 2 2 2 2 2]
	Output MF Type	Linear
	FIS Generation Method	Grid Partitioning
	Learning Algorithm	Hybrid
	Number of Epochs	150

.

The best-performing MLP model was obtained using the Levenberg–Marquardt (L-M) training with a data split of 70% for training, 5% for validation, and 25% for testing. For ANFIS modeling, the best configuration employed the hybrid learning algorithm with gbell membership functions, using a 75% training, 10% validation, and 15% testing split.

The estimation results, showcasing the highest prediction performance for the training and testing, are visually presented in Figure 2. The RMSE for MLP modeling, utilizing 70% training data, is approximately 0.0995 and it is 0.0923 for the testing data. And the R² was 0.8069. It is important to note that, although the test performance curve is displayed alongside the training and validation curves in Figure 2, the test data was not used during training or model selection. In the Neural Network Fitting App of MATLAB R2017b (9.3.0.713579), the model is trained using only the training set, while the validation set is used to prevent overfitting via early stopping. The test set is evaluated only once after training is completed, and the resulting performance metrics and plots are generated for visualization purposes only. Therefore, the independence of the test set was strictly preserved throughout the modeling process.

The estimation results, showcasing the highest prediction performance for the training and testing, are visually presented in Figure 3. For ANFIS modeling with 75% training data, the RMSE was approximately 0.0830 for the training set and it was around 0.1633 for the testing set. The R² values varied across datasets, with about 0.859 for training, 0.4474 for validation, and 0.5053 for testing.

The comparative evaluation of MLP and ANFIS models across different data split configurations is presented in

Table 4. Comparison of MLP and ANFIS performance based on RMSE and R² values computed on the test datasets for different training–validation–testing splits.

Data Splitting Percentage	MLP		ANFIS
	RMSE	R2	RMSE	R2
70% Train	0.0995	0.8048	0.0835	0.8565
5% Validation	0.1033	0.7903	0.2026	0.4796
25% Test	0.0923	0.8069	0.2254	0.3192
70% Train	0.0959	0.8117	0.0835	0.8565
10% Validation	0.0968	0.7916	0.1887	0.5492
20% Test	0.0999	0.8041	0.2386	0.2604
75% Train	0.0933	0.8241	0.0830	0.8588
5% Validation	0.1186	0.7093	0.2027	0.5100
20% Test	0.1137	0.7306	0.2198	0.2791
75% Train	0.0999	0.7959	0.0830	0.8588
10% Validation	0.1026	0.7892	0.2304	0.4474
15% Test	0.1059	0.7634	0.1633	0.5053
80% Train	0.0955	0.8186	0.0852	0.8499
5% Validation	0.1000	0.7732	0.2513	0.4167
15% Test	0.1037	0.7725	0.0702	0.6388
80% Train	0.0971	0.8067	0.0852	0.8499
10% Validation	0.1051	0.7840	0.2440	0.4460
10% Test	0.1048	0.7780	0.1936	0.5555
85% Train	0.0997	0.7947	0.0843	0.8552
5% Validation	0.0932	0.8247	0.2539	0.4391
10% Test	0.1129	0.7695	0.2659	0.1021
85% Train	0.1015	0.7931	0.0843	0.8552
10% Validation	0.0967	0.7953	0.2864	0.3790
5% Test	0.1026	0.7505	0.2406	0.4517

, revealing distinct performance patterns. During the training phase, ANFIS achieved marginally lower RMSE and higher R² values than MLP, suggesting strong fitting capacity. However, MLP demonstrated superior generalization in the testing phase, with lower RMSE and higher R² values compared to ANFIS. For instance, in the 70:5:25 split, MLP yielded an RMSE of 0.0923 and R² of 0.8069 on the test set, whereas ANFIS showed a notable decline with an RMSE of 0.1633 and R² of 0.5053 in the 75:10:15 split. Similar patterns were observed across other testing ratios, confirming that, although ANFIS is effective in capturing training patterns, MLP provides more robust and reliable predictions on unseen data, making it a more suitable model for operational evaporation forecasting.

4. Discussion

This study assessed and compared the performance of Artificial Neural Networks (ANNs) in the form of a Multilayer Perceptron (MLP) and Adaptive Neuro-Fuzzy Inference System (ANFIS) for daily reservoir evaporation estimation using long-term hydrometeorological data from the Atatürk Dam.

The findings regarding the performance of Multilayer Perceptron (MLP) and Adaptive Neuro-Fuzzy Inference System (ANFIS) models in predicting daily reservoir evaporation across varying training–test splits demonstrate that both models possess considerable predictive capabilities. Nonetheless, MLP consistently yields more robust and generalizable outcomes during the testing phase, whereas ANFIS exhibits a pronounced tendency toward overfitting.

The most balanced configuration for the MLP model was achieved with a data split of 70% for training, 5% for validation, and 25% for testing, yielding R² values of nearly 0.80 in both the training and testing phases. In contrast, while ANFIS exhibited its highest training performance at a 75% for training, 10% for validation, and 15% for testing ratio (R² = 0.8588), the corresponding test R² declined to 0.5053, highlighting the model’s sensitivity to data partitioning. During the training phase, the Adaptive Neuro-Fuzzy Inference System (ANFIS) consistently outperformed the Multilayer Perceptron (MLP) in terms of lower Root Mean Square Error (RMSE) and higher coefficient of determination (R²) values. For example, with a 75% training split, ANFIS achieved an RMSE of 0.08300 and an R² of 0.85876, whereas MLP, under identical conditions, recorded an RMSE of 0.09993 and an R² of 0.79592. These results indicate that ANFIS is highly effective at identifying localized patterns within the training data. However, this strength also reflects a strong tendency toward overfitting, limiting its ability to generalize to unseen data. Adnan et al. [34] revealed that various ANFIS variants outperformed ANNs, while the studies by Salih et al. [35] and Allawi et al. [36] confirmed the high predictive capacity of ANFIS. However, this proficiency also entails a heightened susceptibility to overfitting. Although ANFIS attained high predictive accuracy across all training configurations, a marked deterioration in test performance was consistently observed. This trend has been corroborated by recent studies (e.g., [15,37,38], which highlight ANFIS’s propensity to overlearn noise and local fluctuations in the training data, ultimately compromising its generalization performance on unseen datasets. Zhou et al. [38] also confirmed that ANNs with ensemble feature selection yielded superior evapotranspiration predictions across heterogeneous climatic zones.

MLP models demonstrated more balanced and stable performance across both training and test datasets, exhibiting high accuracy even under lower training ratios. For instance, with a 70% training split, MLP achieved an RMSE of 0.0923 and an R² of 0.8069 in testing, outperforming ANFIS, which under the same conditions yielded an RMSE of 0.2254 and an R² of 0.3192. Furthermore, MLP models can be efficiently trained using algorithms such as Levenberg–Marquardt. These findings are consistent with those of Arya-Azar et al. [17], who attribute the superior generalization ability of ANN architectures to their layered nonlinear transformation capacity and flexible weight optimization mechanisms.

The ability of ANNs to deliver consistent performance, particularly during the testing phase, is of critical importance for water management systems, where reliable predictions under variable meteorological conditions are essential. Evaporation processes are highly sensitive to variations in temperature, humidity, solar radiation, and wind; therefore, models such as Artificial Neural Networks (ANNs), which can effectively adapt to such fluctuations, are becoming increasingly critical for long-term water resource planning. Bouramtane et al. [39] likewise emphasized that models capable of maintaining stability amid fluctuating input parameters are indispensable for operational forecasting systems.

Recent advances in hydrology and water resource modeling have increasingly highlighted the potential of Artificial Neural Networks (ANN) and the Adaptive Neuro-Fuzzy Inference System (ANFIS) in evaporation prediction. However, it is widely acknowledged that the success of these models is strongly dependent on the choice of input features and model configuration. Khosravi et al. [40] demonstrated this by comparing the predictive accuracy of eight Machine Learning (ML) and Deep Learning (DL) algorithms using 30 years of meteorological data in Iran. Their results showed that hybrid approaches consistently outperformed individual models.

The importance of input feature selection and model structure is also well-documented. For example, Marouane et al. [19] reported that including water temperature improved the performance of ANFIS-HHO models in Algeria. This is consistent with our ReliefF-based variable selection, which identified key predictors and significantly enhanced model performance. Likewise, Adnan Ikram et al. [41] integrated ANFIS with the Whale Optimization Algorithm (WOA) for monthly pan evaporation estimation in the Dongting Lake basin. Their findings emphasized the dominant role of maximum temperature and the added value of periodicity inputs, even under data-scarce conditions.

In mountainous regions such as the Indian Himalayas, Malik et al. [42] demonstrated that a hybrid neuro-fuzzy system (CANFIS) outperformed conventional ANN models for monthly evaporation simulation. These results collectively underscore that integrating neuro-fuzzy logic with optimization algorithms offers a powerful and adaptive framework for modeling complex hydrological processes.

In parallel, ensemble models have gained attention for their enhanced robustness and generalizability. Ehteram et al. [43] introduced a composite ANN framework that consistently outperformed individual hybrid models across different stations, reinforcing the value of ensemble learning. Similarly, our ANN-based model maintained stable performance across all data splits, supporting its reliability for operational forecasting.

Despite these advances, model transferability remains a key limitation. Alsumaiei et al. [44] found that ANN models trained with limited inputs (e.g., temperature and wind) performed adequately at local scales but failed to capture extreme evaporation events in arid environments. Our findings corroborate this concern, suggesting that AI-based models require regional calibration to preserve accuracy under diverse climatic regimes.

In addition to the comparison between ANN and ANFIS, it is essential to contextualize the findings against other widely used evaporation prediction models. Recent studies have shown that ensemble learning and deep learning techniques have emerged as powerful alternatives to traditional AI approaches in evaporation modeling. For instance, Abed et al. (2022) [45] reported that Random Forest (RF) outperformed ANNs in long-term pan evaporation prediction, primarily due to its inherent capability to rank variable importance and handle complex interactions without overfitting. Likewise, Bilali et al. (2022) [10] demonstrated that Extreme Gradient Boosting (XGBoost) yielded high predictive accuracy (NSE from 0.76–0.83) and interpretability through integration with SHAP and LIME tools, which facilitated identification of dominant climatic drivers such as solar radiation and air temperature.

Furthermore, Mehra et al. (2025) [46] applied Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks for evaporation modeling and showed that these deep learning architectures outperformed both classical machine learning methods and traditional ANNs, especially in capturing temporal dependencies in time series data. In a recent study, deep LSTM architectures alone have shown improved accuracy over traditional ANN and empirical models for daily pan evaporation estimation, especially with limited input features. Similarly, CNN models have demonstrated high accuracy in capturing the nonlinearities of evaporation processes, often outperforming other machine learning frameworks [45]. Additionally, hybrid and optimization-enhanced frameworks such as ANFIS coupled with Particle Swarm Optimization (PSO) or the Whale Optimization Algorithm (WOA) have demonstrated superior predictive capabilities across various hydroclimatic regions [19,41].

Compared to these models, our findings suggest that, while Artificial Neural Networks (ANNs) in the form of a Multilayer Perceptron (MLP) exhibited superior generalization performance over ANFIS in this study, the core architecture of many recent high-performing models still builds upon MLP or ANFIS structures—whether through optimization, hybridization, or deep expansion. This reinforces the practical value of our direct comparative analysis and highlights that enhancing and tuning these foundational architectures continues to be a central strategy in evaporation prediction modeling. Future work could extend the current comparative framework by including RF, XGBoost, and other emerging models such as deep learning architectures (e.g., LSTM), which have also shown promise in evaporation and evapotranspiration modeling under variable climate conditions.

Taken together, the findings from Table 4 and recent literature indicate that the MLP model provides a more stable and generalizable approach for evaporation prediction, maintaining its effectiveness especially in large-scale and real-time applications. In contrast, although ANFIS models can achieve high accuracy on training data, they require additional support from optimization algorithms (e.g., PSO, GA, SFLA) or pruning techniques to improve their generalization performance and prevent overfitting. Ultimately, model selection should be guided by data volume, preprocessing level, and application objectives, with MLP being the preferred choice for general-purpose and operational use.

5. Conclusions

The findings highlight the strategic potential of AI-based models in supporting more adaptive and climate-resilient water resource management. Combining the interpretability of ANFIS with the adaptability of MLP or enriching these models with evolutionary algorithms and feature selection techniques such as ReliefF can yield high-performance and flexible systems for evaporation prediction. When integrated with decision support tools, such systems can contribute to mitigating the impacts of drought, managing reservoir operations, optimizing water distribution plans, formulating climate-responsive policies, and ultimately supporting inclusive and sustainable water resource management.

Moreover, the implementation of these models within early warning and monitoring systems can enhance preparedness against hydrological extremes. They also enable real-time updates of operational rules for reservoirs, allowing for more responsive and efficient water allocation.

To further strengthen the applicability of such models, future research should explore regional calibration, integration with climate projections, and real-time data assimilation techniques.