Prediction of groundwater level in shallow and complex aquifers using hybrid soft computing models and metaheuristic algorithms Authors: Saeed Mohammadpour, Arash Ebrahimabadi, Hamidreza Rabiefar Published: 2026 ABSTRACT Predicting groundwater level dynamics in shallow, unconfined aquifers remain a persistent challenge due to their high sensitivity to heterogeneous recharge, rapid anthropogenic extraction, and nonlinear interactions with climatic drivers conditions under which conventional models often fail to generalize. Here, we introduce two novel hybrid frameworks ANN-HBO and ANN-POA that couple artificial neural networks with metaheuristic optimization algorithms (Honey Badger Optimization and Pelican Optimization) to dynamically optimize network weights in data-scarce, non-stationary environments where deep learning architectures exhibit poor convergence and overfitting. The study is conducted in the Astaneh-Kuchesfahan aquifer, northern Iran, a shallow alluvial system limited storage capacity, and strong seasonal fluctuations driven by irrigation demand and evaporation. Hydraulic conductivity varies spatially from 10 to 35 m/day, reflecting moderate to high transmissivity (approximately 200–600 m2/day), typical of well-drained fluvial systems. Using 256 months of hydro-meteorological records, we identify the most informative predictors lagged groundwater levels (1–3 months), temperature, precipitation, evaporation, and abstraction rates via the Minimum Redundancy Maximum Relevance (MRMR) algorithm, generating four optimized input scenarios. Compared to a standalone ANN, ANN-POA reduces RMSE by 29.3% and MAE by 24.1% on average across three observation wells, achieving test RMSE values of 0.213–0.287 m and MAE of 0.260–0.437 m. Lagged groundwater level (1-month) and monthly precipitation are identified as the dominant predictors, confirming the system’s strong dependence on recent recharge. The superior robustness of ANN-POA over ANN-HBO and baseline models demonstrates its efficacy in navigating complex, high-dimensional parameter spaces without requiring extensive training data. This approach enables water managers in data-limited regions to transition from reactive to anticipatory aquifer management: identifying early signs of depletion, prioritizing managed recharge, and calibrating extraction quotas based on dynamic hydrological responses rather than static averages. Our framework offers a scalable, physics-informed, and low-cost solution for sustainable groundwater governance in vulnerable shallow aquifers worldwide.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-43005-w. Introduction Groundwater is considered one of the most valuable and vital water resources for various uses. excessive groundwater withdrawal and inadequate planning in its allocation and consumption can have undeniable negative consequences on the environment and water resources management1. Consequently, groundwater has attained a crucial and prominent position as a valuable and renewable resource within the management and planning cycle of water resources policy, both quantitatively and qualitatively. Due to increased demand and consumption driven by climatic and anthropogenic changes, it has captured the special focus of governmental and sovereign institutions, as well as local and state governments, becoming a central point of attention2. Understanding the nature of a system and its interconnected and integrated components, as well as how changes in one component affect other parts of the system, remains a fundamental challenge in the field of water resources. In this regard, quantifying groundwater responses gains particular importance. Therefore, achieving an appropriate prediction necessitates understanding the relationships governing climatic drivers, comprehending the hydrological nature of groundwater, and ensuring its optimal use in various sectors such as agriculture, industries and services, and other related areas3. The accurate simulation of a physical process demands a thorough understanding of its governing relationships and intrinsic complexities. This often leads to significant challenges in the modeling and calibration process, given the extensive agricultural, hydrological, and geohydrological data required, ultimately impeding the development of effective models. Moreover, achieving a unified prediction of groundwater levels by integrating the responses of various interacting factors proves to be a particularly challenging task4. Controlling and monitoring groundwater level (GWL) can be of particular interest to groundwater specialists for short-term and long-term planning in water resources management5. In this regard, the use of machine learning models will be highly practical for non-linear models capable of adequately predicting groundwater level changes, especially when the modeling process in such models is performed without deep and sufficient knowledge of the physical parameters of the problem [Reference]. These models will be applied based on historical data records and the identification of suitable patterns for forecasting future scenarios6. ML algorithms possess the ability to identify complex structures without considering the nature and physics of the problem, enabling accurate predictions for irregular data. Therefore, the advantage of these methods over physical, traditional, and other conventional approaches lies in ML models’ capacity for simulating processes and achieving accurate and consistent predictions without requiring deep knowledge of climatic parameters7. Evolutionary algorithms have accelerated scientific discoveries by significantly enhancing efficiency, accuracy, and speed in solving complex problems, and by enabling the analysis of massive datasets and the discovery of hidden patterns. Furthermore, hybrid models, by integrating multiple computational approaches, have become a powerful strategy for overcoming the limitations of individual algorithm8. Optimizer models (metaheuristics) help Artificial Neural Networks to perform the weight and bias optimization process with greater precision, ultimately leading to improved prediction accuracy9. Aquifers whose groundwater level is close to the surface, although considered an important water source, also come with difficulties. These aquifers are easily replenished by precipitation, but this very proximity of water to the surface creates serious problems. Among these issues are water loss through evapotranspiration, salinization of agricultural lands, and soil waterlogging10. Given these challenges, proper and integrated management of these aquifers is essential. This management should include strategies such as establishing effective drainage systems, optimizing irrigation methods, improving soil quality and structure, and continuously monitoring water quality. Ultimately, considering climate change and increasing water demand, the intelligent management of these aquifers is of high importance for environmental preservation and water supply for human use and agriculture, requiring close cooperation between executive and local authorities11,12. Nafouanti et al.13(2025) used deep learning and hybrid models to predict groundwater levels and their connection to land subsidence and water quality in the North China Plain. Their LSTM model accurately forecasted groundwater levels, while a hybrid Random Forest-KNN model proved most effective for predicting fluoride contamination. Tahkur et al.14 a hybrid PCA–DWT–LSTM model for groundwater level prediction was developed. Their results showed that this model was significantly more accurate than a model without the wavelet transform (DWT) component. Zhu et al.15 developed a Bayesian ensemble framework, using a stacking algorithm for probabilistic groundwater level forecasting. Their ensemble model was more accurate and provided better uncertainty estimates than any of the individual machine learning models. Of course. Here is the literature review entry. Igwebuike et al.16 compared two machine learning algorithms (Random Forest and Support Vector Machine) with two deep learning algorithms (SimpleRNN and LSTM). Their analysis revealed that the machine learning models outperformed the deep learning models. The Support Vector Machine (SVM) was identified as the best overall model, achieving the highest accuracy with the lowest RMSE. Most research conducted to date in this field pertains to aquifers with deeper water tables or in arid climates where groundwater levels exhibit a clear declining pattern or a distinct sinusoidal pattern. However, accurate prediction of groundwater levels in shallow and complex aquifers is essential, and its implementation is unavoidable. Because these aquifers are highly susceptible to climatic changes, evapotranspiration, and soil salinization, without a proper understanding of water table fluctuations, these vital resources which are of critical and strategic importance for various uses such as agriculture and drinking water cannot be managed correctly. Thakur and Karmakar17 (2025) conducted a comparative analysis of ANN, LSTM, and hybrid PSO-LSTM algorithms for groundwater level prediction. Evaluation of the hybrid PSO-LSTM model showed a significant improvement compared to the standalone LSTM. Furthermore, the results indicate that the implemented hybrid PSO-LSTM algorithm outperforms both the traditional ANN and the standard LSTM. Mirboluki et al.18 (2024) addressed groundwater level (GWL) estimation using enhanced deep learning and machine learning methods. The study compared the Grey Wolf Optimizer-Long Short-Term Memory (LSTM-GWO) deep learning model with the Grey Wolf Optimizer-Artificial Neural Network (ANN-GWO) hybrid machine learning model. The results demonstrated that the LSTM-GWO method showed better capability in estimating GWL than the ANN-GWO and ANN methods. Other findings also indicated that the hybrid LSTM-GWO model improved the GWL estimations by approximately 23%. Osman et al.19 (2024) evaluated machine learning algorithms for predicting the impact of climate change on groundwater levels to protect against aquifer depletion. The study compared the results of the Long Short-Term Memory (LSTM), Extreme Gradient Boosting (XGBoost), Artificial Neural Network (ANN), and Support Vector Regression (SVR) models. Their findings indicated that the XGBoost model achieved the highest predictive performance (RMSE = 0.026) followed by LSTM (RMSE = 0.027, 0.050, 0.085) respectively. Yeganeh et al.20 (2024) compared deep learning models and shallow models for groundwater level prediction, focusing on the integration of multi-piezometric data. Their findings showed that the Long Short-Term Memory (LSTM) model significantly outperformed its competitors. Its superiority over the next best technique was approximately 13%. Abdi et al.21 (2024) utilized interpolation techniques in deep learning models to enhance groundwater level (GWL) prediction accuracy. The study was conducted on the Azarshahr plain aquifer in Iran and involved 34 observation wells. Three interpolation methods Kriging, Support Vector Machine (SVM), and M5P were employed. Results showed that the M5P method was selected as the most accurate method (RMSE} = 1.83). Furthermore, a comparison of the Deep Neural Network (DNN), Convolutional Neural Network (CNN),and LSTM models indicated that the CNN model’s performance was more favorable. Other findings also revealed that improving data quality through interpolation significantly increased prediction accuracy by approximately 90%. Chang et al.22 (2025) addressed advanced groundwater level (GWL) prediction using a hybrid deep learning model. The study utilized a hybrid deep learning model (CNN-BP) and a machine learning model based on Back-Propagation Neural Networks (BPNN) to predict groundwater levels in the Zhoushi River alluvial fan. The findings indicated that the CNN-BP model significantly outperformed the BPNN model, achieving (R2 values between 0.94 and 0.98) at various stations, and effectively mitigated time lag issues. The Astaneh-Kuchesfahan aquifer, represents a prominent example of a shallow and hydrogeologically complex aquifer in coastal and alluvial plains, shaped by the interplay of geological, climatic, and anthropogenic factors. In this region, the groundwater level (GWL) remains persistently close to the land surface a condition that not only promotes waterlogging and soil salinization but also elevates the role of evaporation as a dominant, nonlinear driver of GWL dynamics. The proximity of the water table to the surface induces a strong and highly nonlinear coupling between GWL and evapotranspiration. This direct coupling, particularly intensified by high evaporation rates from the capillary fringe even during periods of seasonal waterlogging constitutes the primary challenge in accurately modeling GWL behavior in such systems. Conventional modeling approaches, which often treat evaporation as a boundary condition or linear forcing term, fail to capture this dynamic feedback, leading to significant predictive inaccuracies. Although Artificial Intelligence-based models, particularly Artificial Neural Networks (ANNs), have achieved success in groundwater level prediction, their performance is systematically affected by fundamental limitations. These inherent limitations include a strong reliance on local optimization algorithms that easily become trapped in local minima within high-dimensional and non-linear parameter spaces, and an inability to optimally tune complex model structures under conditions where input data is characterized by high noise, uncertainty, or severe non-linear relationships (such as the connection between evaporation and water level in shallow aquifers). Therefore, integrating ANN with metaheuristic algorithms can mitigate these limitations; however, most previous works have utilized classical algorithms, which themselves face challenges including premature convergence, the need for manual parameter tuning, and inefficiency in complex search spaces. In this regard, two novel metaheuristic algorithms, the Honey Badger Optimization (HBO) and the Pelican Optimization Algorithm (POA), were selected to optimize the ANN for the complex conditions of the Astaneh-Kuchesfahan aquifer due to their unique structural features. Although advanced models such as Bayesian networks, ensemble methods (e.g., XGBoost, Random Forest), and deep learning models (e.g., LSTM, CNN, RNN) networks have been successfully applied to groundwater level (GWL) prediction, they exhibit fundamental limitations under the specific conditions of shallow aquifers with high evaporation influence. Bayesian networks operate under linear or quasi-linear probabilistic assumptions and are incapable of capturing the complex nonlinear interactions between evaporation and GWL dynamics, particularly during waterlogging events. Moreover, their performance is highly sensitive to large training datasets and structural priors, which are often unavailable in data-scarce regions such as the Astaneh-Kuchesfahan aquifer, leading to elevated predictive uncertainty. Ensemble tree-based models, while effective in feature selection, struggle to generalize continuous, physically coherent relationships such as the gradual response of GWL to sustained evaporation fluxes due to their discrete, piecewise nature. LSTM models, despite their capacity to learn temporal dependencies, require high-temporal-resolution data (daily or hourly) to perform reliably; when applied to monthly, noisy datasets as in this study they frequently suffer from overfitting or convergence failure. In contrast, the ANN-HBO and ANN-POA models overcome these limitations by synergistically combining the nonlinear approximation capability of ANNs with the global optimization power of the Honey Badger Optimization (HBO) and Pelican Optimization Algorithm (POA). These metaheuristics enable robust parameter tuning without requiring high-frequency data or structural assumptions, effectively navigating high-dimensional, noisy search spaces to minimize model sensitivity to uncertainty and improve predictive stability. This study develops two novel hybrid models ANN-HBO and ANN-POA to address the limitations of conventional and deep learning approaches in predicting groundwater levels in shallow aquifers under strong evaporation influence, where data are sparse, noisy, and characterized by complex nonlinear dynamics. By leveraging global optimization through the Honey Badger Optimization (HBO) and Pelican Optimization Algorithm (POA) to tune ANN parameters, these models enhance prediction accuracy, stability, and generalizability in conditions where traditional models fail due to their reliance on high-frequency data, linear assumptions, or tree-based structures incapable of capturing continuous physical feedbacks. Review of groundwater prediction using machine learning models Machine learning can examine vast amounts of data and uncover specific trends and patterns that are unknown to humans. These models are capable of analyzing very large datasets and providing insights that aid in the decision-making process. Given the various limitations in using physical and conceptual models, the application of machine learning models has seen significant growth among researchers across diverse fields. Numerous studies have been presented for predicting groundwater levels using standalone and enhanced hybrid models, with Table 1 summarizing various groundwater level prediction studies utilizing machine learning models. This table contains information on authors, study area, models used (standalone and hybrid), data time series extent (statistical years), temporal step scale (hourly, daily, and monthly), and input parameters. The parameters are, respectively, Precipitation (R), Temperature (T), Humidity (H), Relative Humidity (RH), Groundwater Level (GWL), Potential Evapotranspiration (EVP), Flow Discharge (Q), Minimum Temperature (Tmin), Maximum Temperature (Tmax), and Mean/Average Temperature (Tmean or Tave).Table 1Overview of GWL prediction studies utilizing machine learning models.NoAuthor (year)Study areaMachin LearningData SpanTime stepInput data1(Adamowski and Chan, 2011)5CanadaWA-ANN, ANN, ARIMA1983–2009MonthlyP, E, GWL2(Shiri et al., 2013)23KoreaANFIS, FFNN, GEP, SVM, ARMA2001–2008DailyET, P3(Djurovic et al., 2015)24SerbiaANFIS, FFNN1990–2010DailyP, EP, Tmax, Tmin4(Nair and Sindhu, 2016)25IndiaMLP2002–2016MonthlyR, P, T, H5(Nourani and Mousavi, 2016)26IranANFIS, FFNN, WANFIS WFFNN2001–2011MonthlyP, Q6(Gong et al., 2016)27USAANFIS, FFNN, SVM1998–2009MonthlyGWL, P, LL, Tmean7(Yoon et al., 2016)28South KoreaANN, SVM2003–2008DailyGWL, P8(Hong, 2017)29TaiwanFFBPNN2008HourlyGWL9(Yadav et al., 2017)30CanadaELM2006–2014MonthlyGWL, R, T, EVP10(Zare and Koch, 2018)31IranANFIS-FCM, WANFIS-FCM1991–2013MonthlyP, GWL11(Wunsch, Liesch and Broda, 2018)32GermanyNARX, ANN1948–2008DailyP, T12(Lee, Lee and Yoon, 2019)33South KoreaFFBPNN2014DailyRH, T, R13(Bak and Bae, 2019)34KoreaANFIS2015–2017DailyP, Tmean14(Jahanara and Khodashenas, 2019)35USANF-GMDH-PSO, RBFNN-NF-GMDH-PSO2001–2008DailyGWL15(Guzman et al., 2019)36USANARX-ANN, RBF-SVR1985–1994DailyGWL, P, ET16(Derbela and Nouiri, 2020)37TunisiaANN2000–2018MonthlyR, GW, P17(Al-Jami et al., 2020)38BangladeshNARX, ANN1980–2013MonthlyP, T, H18(Sun et al., 2022)8ChinaARIMA, LSTM, BP-ANN1956–2002DailyGWL, P,T19(Mirzania et al., 2024)39IranWHO–ANN, ESOA–ANN1992–2015MonthlyGWL20(Sahoo and Satapathy, 2025)40IndiaSVM-SOA, ANN-PSO2002–2023MonthlyGWL, P,T Table 1Overview of GWL prediction studies utilizing machine learning models.NoAuthor (year)Study areaMachin LearningData SpanTime stepInput data1(Adamowski and Chan, 2011)5CanadaWA-ANN, ANN, ARIMA1983–2009MonthlyP, E, GWL2(Shiri et al., 2013)23KoreaANFIS, FFNN, GEP, SVM, ARMA2001–2008DailyET, P3(Djurovic et al., 2015)24SerbiaANFIS, FFNN1990–2010DailyP, EP, Tmax, Tmin4(Nair and Sindhu, 2016)25IndiaMLP2002–2016MonthlyR, P, T, H5(Nourani and Mousavi, 2016)26IranANFIS, FFNN, WANFIS WFFNN2001–2011MonthlyP, Q6(Gong et al., 2016)27USAANFIS, FFNN, SVM1998–2009MonthlyGWL, P, LL, Tmean7(Yoon et al., 2016)28South KoreaANN, SVM2003–2008DailyGWL, P8(Hong, 2017)29TaiwanFFBPNN2008HourlyGWL9(Yadav et al., 2017)30CanadaELM2006–2014MonthlyGWL, R, T, EVP10(Zare and Koch, 2018)31IranANFIS-FCM, WANFIS-FCM1991–2013MonthlyP, GWL11(Wunsch, Liesch and Broda, 2018)32GermanyNARX, ANN1948–2008DailyP, T12(Lee, Lee and Yoon, 2019)33South KoreaFFBPNN2014DailyRH, T, R13(Bak and Bae, 2019)34KoreaANFIS2015–2017DailyP, Tmean14(Jahanara and Khodashenas, 2019)35USANF-GMDH-PSO, RBFNN-NF-GMDH-PSO2001–2008DailyGWL15(Guzman et al., 2019)36USANARX-ANN, RBF-SVR1985–1994DailyGWL, P, ET16(Derbela and Nouiri, 2020)37TunisiaANN2000–2018MonthlyR, GW, P17(Al-Jami et al., 2020)38BangladeshNARX, ANN1980–2013MonthlyP, T, H18(Sun et al., 2022)8ChinaARIMA, LSTM, BP-ANN1956–2002DailyGWL, P,T19(Mirzania et al., 2024)39IranWHO–ANN, ESOA–ANN1992–2015MonthlyGWL20(Sahoo and Satapathy, 2025)40IndiaSVM-SOA, ANN-PSO2002–2023MonthlyGWL, P,T Overview of GWL prediction studies utilizing machine learning models. Investigated region Northern Iran is a region where precipitation is significantly higher than the Iranian average, and consequently, groundwater in this area typically exhibits a different behavior compared to other aquifers in the country. Due to adequate recharge and the soil’s nature, the groundwater table is predominantly close to the surface, and the groundwater level follows an irregular pattern. The Astaneh-Kuchesfahan plain aquifer is one of the important water supply sources for Gilan Province, formed on the alluvial fan of the Sefidrud River, with an area of approximately 1100 square kilometers. This area is located between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${40}^{0}{12}{\prime}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${50}^{0}{5}{\prime}$$\end{document} eastern longitude and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${37}^{0}{7}{\prime}$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${37}^{0}{25}{\prime}$$\end{document} northern latitude. This region is considered among the high-rainfall areas of Iran, such that its average annual precipitation, based on 40 years of statistics, is about 1430 mm. This region is considered among the high-rainfall areas of Iran, such that its average annual precipitation, based on 40 years of statistics, is about 1430 mm. The abstraction rate from the region’s groundwater resources is approximately 2.44 million cubic meters per year. The most important river in this basin is the Sefidrud, with a length of 670 km. The average annual discharge of this river has been 3998.4 million cubic meters41. The Astaneh-Kuchesfahan plain has an unconfined aquifer and rests on an impermeable Jurassic layer. The maximum thickness of the aquifer is 250 m in the center of the plain, and its minimum thickness is 40 m in the north of the plain. Figure 1 shows the geographical location of the Astaneh-Kuchesfahan aquifer. Also Fig. 2 shows geographical location of the studied aquifer. Although the Astaneh-Kuchesfahan aquifer is often labeled as having “high potential” due to its shallow water table and moderate recharge, this reflects socio-hydrological accessibility, not resource abundance. Its shallow depth enables low-cost extraction via shallow wells, making it the dominant source of water for domestic and agricultural use in rural Gilan, where groundwater satisfies over 85% of irrigation demand. However, this same accessibility coupled with high evaporation, poor drainage, and intensive seasonal pumping renders the system highly susceptible to degradation, including waterlogging, salinization, and quality decline. This study focuses on shallow groundwater not because it is abundant, but because it is critically vulnerable: accurate prediction is essential for sustainable management in this high-stress, data-limited environment.The Astaneh-Kuchesfahan aquifer is an unconfined, shallow alluvial system with a water table typically less than 5 m below ground surface, characteristic of coastal plains in Gilan Province. No significant confined aquifer units exist within the immediate study area; deeper aquifers (> 50 m) are hydraulically disconnected from the shallow system due to the presence of low-permeability clay layers, as confirmed by regional hydrogeological cross-sections provided by the Gilan Regional Water Company. Consequently, groundwater abstraction occurs exclusively from the unconfined aquifer, primarily for agricultural irrigation. The average annual abstraction volume over the 2005–2022 period is estimated at approximately 42 million cubic meters (MCM), corresponding to a mean monthly abstraction rate of ~ 3.5 MCM/month, derived from official well licensing records and pumping hour estimates. This extraction rate, combined with intense evaporation and poor natural drainage, exerts significant stress on the shallow aquifer, contributing to waterlogging in wet seasons and salinity buildup during dry periods. Given the absence of vertical connectivity between shallow and deep aquifers, all modeling efforts in this study focus solely on the unconfined system, with abstraction rate included as a dynamic input variable in the prediction models. The spatial extent and water table dynamics of the unconfined Astaneh-Kuchesfahan aquifer were defined based on data from 38 active piezometric wells distributed across the study area, as maintained by the Gilan Regional Water Company. These wells are strategically located to capture groundwater fluctuations in both agricultural and residential zones, with an average spacing of approximately 1.5 km, ensuring adequate spatial coverage for regional-scale modeling. All wells are completed within the shallow alluvial deposits (depths ranging from 8 to 25 m) and screened in the saturated zone of the unconfined aquifer. Monthly groundwater level measurements from these wells were used to represent spatiotemporal variations in the aquifer system and to support the calibration and validation of the proposed models. The modeling framework incorporates both meteorological and anthropogenic inputs to capture the key drivers of groundwater level fluctuations in the Astaneh-Kuchesfahan aquifer. Meteorological variables include monthly precipitation, evaporation, and temperature, obtained from the Iranian Meteorological Organization. Human-induced factors are represented by groundwater abstraction rates (in MCM/month), derived from well licensing records and pumping hour logs provided by the Gilan Regional Water Company. All data span a continuous 256-month period from January 2000 to December 2022, ensuring sufficient temporal coverage to account for interannual climate variability and long-term trends in water use. These inputs were used as predictors in both standalone and hybrid ANN models, with all series preprocessed for consistency and stationarity prior to model training. Figure 3 show the spatial distribution and location of the wells.Fig. 1Geographical location of Astaneh-Kuchesfahan aquifer. The map was generated by the authors using ArcGIS 10.8.2 (https://support.esri.com/en-us/patches-updates/2024/arcgis-desktop-engine-10-8-2-general-component-updates-patch-2) and does not require any permissions.Fig. 2Geographical location of the studied aquifer. The map was generated by the Khonok et al., 2022. (https://doi.org/10.1007/s13762-021–03421-z) using ArcGIS 10.2 (https://support.esri.com/en/download/2093).Fig. 3The spatial distribution and location of the wells. The map was generated by the authors using ArcGIS 10.8.2 (https://support.esri.com/en-us/patches-updates/2024/arcgis-desktop-engine-10-8-2-general-component-updates-patch-2) and does not require any permissions. Geographical location of Astaneh-Kuchesfahan aquifer. The map was generated by the authors using ArcGIS 10.8.2 (https://support.esri.com/en-us/patches-updates/2024/arcgis-desktop-engine-10-8-2-general-component-updates-patch-2) and does not require any permissions. Geographical location of the studied aquifer. The map was generated by the Khonok et al., 2022. (https://doi.org/10.1007/s13762-021–03421-z) using ArcGIS 10.2 (https://support.esri.com/en/download/2093). The spatial distribution and location of the wells. The map was generated by the authors using ArcGIS 10.8.2 (https://support.esri.com/en-us/patches-updates/2024/arcgis-desktop-engine-10-8-2-general-component-updates-patch-2) and does not require any permissions. Geology of the Astaneh-Kuchesfahan aquifer Based on the geological profiles of exploratory and piezometric wells drilled in the Astaneh-Kuchesfahan study area, the alluvial aquifer is unconfined and covers an area of 989.51 km2, encompassing the entire plain. The widest part of the Gilan coastal plain in the mentioned area, between the Sefidrud Delta and Imamzadeh Hashem, is approximately 54 km wide (Fig. 4). The Fumanat Heights are located on the left bank (west) of the Sefidrud River, while the Dorfak and Asman Sara Mountains are on the right bank (east). The Caspian Sea borders this area to the north, and the Manjil Dam Lake lies to its south. The aquifer’s constituent material is composed of deltaic and alluvial sediments of the Sefidrud River. In the southern sections, these particles are coarser-grained, creating better quantitative and qualitative conditions. Fine-grained sediments form lenticular clay lenses of varying thickness in different places, complicating the distribution of the confined aquifer. Geologically, the southern part of the aquifer is made up of coarse-grained materials and relatively hard formations, while the northern part of the study area consists of fine-grained materials42. In the aforementioned alluvial aquifer, the elevation of groundwater level contours (equipotential lines) varies between 1 and 10 m. The elevation of groundwater level contours decreases from south to north. In the alluvial fan areas located in the south of the alluvial aquifer, the depth to groundwater is more than 10 m, which decreases northward, reaching its minimum value in the coastal plains of the basin. The total area of groundwater evaporation zones in this alluvial aquifer is generally 921.81 km2, which constitutes 93.21 percent of the total area of the alluvial aquifers. Conversely, in other regions, groundwater depths are less than 1 m, encompassing an area of approximately 17.33 km2 within this alluvial aquifer. The extent of zones with groundwater depths between 1 to 3 m and 3 to 5 m in this alluvial aquifer are 886.51 and 37.97 km2, respectively. Notably, the majority of the aquifer is characterized by a groundwater depth of 2 m. The general trend of groundwater level contours in the aquifer is east–west, and their elevations decrease from south to north. In this aquifer, the elevation of groundwater level contours varies between -25 and 40 m. The maximum groundwater level, with an elevation of 40 m, pertains to the alluvial fan areas located in the south of the alluvial aquifer, and its minimum, at -25 m, is found in the coastal lands located in the north of this aquifer. Generally, across the alluvial aquifer, the convexity of the groundwater level contours is towards the higher elevations, and the groundwater flow lines converge towards the river axis. In the southern lands of the alluvial aquifer, the hydraulic gradient of the groundwater is high, decreasing towards the north43. The groundwater flow direction in this alluvial aquifer is from south to north. The hydraulic gradient of the groundwater is 4 per thousand in the alluvial fan areas, 2 per thousand in the central areas of the aquifer, and less than 1 per thousand in the downstream areas of the aquifer.Fig. 4Geological Map of the Astaneh-Kuchesfahan Aquifer. The map was generated by the authors using ArcGIS 10.8.2 (https://support.esri.com/en-us/patches-updates/2024/arcgis-desktop-engine-10-8-2-general-component-updates-patch-2) and does not require any permissions. Geological Map of the Astaneh-Kuchesfahan Aquifer. The map was generated by the authors using ArcGIS 10.8.2 (https://support.esri.com/en-us/patches-updates/2024/arcgis-desktop-engine-10-8-2-general-component-updates-patch-2) and does not require any permissions. Methodology and research methods Monthly hydroclimatic and groundwater data spanning 256 consecutive months (from 2002–2020) were obtained from the Gilan Regional Water Company and the Iranian Meteorological Organization. Data Preprocessing is a critical stage in any Machine Learning (ML) project, as the quality of the input data directly impacts the performance and accuracy of the resulting model. The objective of this operation is to transform raw data into a clean, consistent, and suitable format for model training. Data preprocessing operations typically encompass four main stages: Data Cleaning (handling deficiencies, errors, and inappropriate data), Data Integration (merging and removing duplicates), Data Reduction (dimensionality reduction, feature selection, feature extraction, and numerosity reduction), and Data Transformation (normalization/scaling and standardization).Correct preprocessing ensures that Machine Learning models focus on learning the real patterns inherent in the data rather than spending time addressing noise and inconsistency. This directly leads to improved Generalization and reduced model error. All input and target variables were scaled to the [0, 1] interval using Min–Max normalization. Data normalization is performed using the following equation:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{norm}=\frac{x-{x}_{min}}{{x}_{max}-{x}_{min}}$$\end{document}where x: denotes the original input feature value (raw data to be transformed), xnorm: is the normalized output value (feature value after scaling), xmin: is the minimum observed value of the feature (smallest value in the feature column or dataset), and xmax is the maximum observed value of the feature (largest value in the feature column or dataset). Minimum Redundancy Maximum Relevance (MRMR) is a popular and widely used feature selection method in machine learning that simultaneously optimizes two opposing criteria: maximum relevance to the target variable and minimum redundancy among selected features. Specifically, the Mutual Information between two random variables, x and y, is mathematically formulated based on their joint and marginal probabilistic density functions:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I(x;y)=p\left(x,y\right)\mathrm{log}\iint \frac{p\left(x,y\right)}{p\left(x\right)p\left(y\right)}dxdy$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Max\left(\mathrm{Relevance}\left(S,c\right)\right),D=\frac{1}{\left|S\right|}\sum_{{x}_{i}\in S} I\left({x}_{i},c\right)$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Min(\mathrm{Redundancy}(S)),R=\frac{1}{|S{|}^{2}}\sum_{{x}_{i},{x}_{j}\in S} I\left({x}_{i},{x}_{j}\right)$$\end{document} The maximum relevance criterion is satisfied by selecting the feature (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}$$\end{document}) that yields the highest mutual information, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I\left({x}_{i};c\right)$$\end{document}, with the target class (c). The MRMR methodology ranks features by concurrently minimizing redundancy among the selected subset and maximizing their relevance to the target variable. This optimization process is driven by the operator Φ :5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$max\Phi \left(D,R\right),\Phi =D-R$$\end{document} That maxΦ: Maximize the Objective Function, Φ: MRMR Objective Function / Operator, D: Maximum Relevance, R: Minimum Redundancy and D – R: Difference Criterion. The study is based on a continuous 256-month observational period (January 2000–December 2020), with data divided into training (70%) and testing (30%) subsets to ensure reliable model evaluation while preserving temporal dependencies; prior to modeling, comprehensive preprocessing was applied, including Min–Max normalization to scale all variables to the [0,1] range, outlier detection and robust replacement via the interquartile range (IQR) method, and linear interpolation for minor missing values (< 2% of observations), all of which were implemented in MATLAB R2023b to ensure computational consistency; the artificial neural network (ANN) was then employed both as a standalone model and hybridized with the Pelican Optimization Algorithm (POA) and Honey Badger Optimization (HBO) to globally optimize its weights and biases, thereby enhancing predictive accuracy under conditions of high nonlinearity and limited data resolution, with the overall methodology including data preprocessing, feature selection, model training, and validation schematically illustrated in Fig. 5.Fig. 5Overall research methodology flowchart. The map was generated by the authors using EdrawMax Ultimate version 15.1.4. (https://edrawmax.wondershare.com/) and does not require any permissions. Overall research methodology flowchart. The map was generated by the authors using EdrawMax Ultimate version 15.1.4. (https://edrawmax.wondershare.com/) and does not require any permissions. Optimal parameters & model blending For optimal scenario selection and model combination, understanding the relationship between variables and their degree of interrelation is of key importance. In this research, T (t), P(t), E (t), W(t), GWL(n-1), GWL(n-2), GWL(n-3), and were used as input variables. To select the optimal model combination and scenarios, it is necessary to calculate the correlation between input and output variables. To this end, correlation was calculated based on the Pearson correlation coefficient and presented in Table 2. Numerous strategies are used in the feature selection and optimization areas to determine optimal scenarios and select the optimum model combination. A crucial component of data preprocessing in machine learning is feature selection, which lowers data dimensionality, lowers model complexity, prevents overfitting, and improves model interpretability. There are several ways to choose the best combination, such as filter-based approaches, the Chi-Squared Test, the Pearson correlation coefficient, MRMR, Mutual Information (MI), and others. Of these, MRMR and PCC have been used more frequently in the field of water resources and have generally grown in popularity. The primary benefit of MRMR, a multidimensional technique, is that it simultaneously takes into account a feature’s non-redundancy with other chosen features and its relevance to the objective. Accordingly, the MRMR method was employed to identify the optimal model combination. Table 3 presents the model combinations and research scenarios.Table 2Scores of influential parameters on GWL using the MRMR algorithm.ParameterT (t)P(t)E(t)W(t)GWL(t−3)GWL(t−2)GWL(t−1)Score0.10750.10670.10070.34960.01130.05810.2454Table 3Optimal model combination and scenario selection based on MRMR.Scenario numberInput variableOutput variable1W(t), GWL(t-1)GWL2GWL(t-1), GWL(t-2), GWL(t-3)GWL3W(t), GWL(t-1), GWL(t-2), GWL(t-3)GWL4W(t), E (t), T (t), P (t), GWL(t-1), GWL(t-2), GWL(t-3)GWL Table 2Scores of influential parameters on GWL using the MRMR algorithm.ParameterT (t)P(t)E(t)W(t)GWL(t−3)GWL(t−2)GWL(t−1)Score0.10750.10670.10070.34960.01130.05810.2454 Scores of influential parameters on GWL using the MRMR algorithm. Table 3Optimal model combination and scenario selection based on MRMR.Scenario numberInput variableOutput variable1W(t), GWL(t-1)GWL2GWL(t-1), GWL(t-2), GWL(t-3)GWL3W(t), GWL(t-1), GWL(t-2), GWL(t-3)GWL4W(t), E (t), T (t), P (t), GWL(t-1), GWL(t-2), GWL(t-3)GWL Optimal model combination and scenario selection based on MRMR. Metaheuristics in research models Neural networks, machine learning models, and metaheuristic optimization algorithms (HBO and POA) were used in this study to analyze and model the data. The main goal was to use these tools’ capabilities to find trends, forecast outcomes, and improve different situations. In order to solve complicated optimization problems and determine the ideal parameters for the neural networks and machine learning models, metaheuristic algorithms were used. These algorithms can find near-optimal solutions for problems with large and complex search spaces; they are frequently inspired by social or natural phenomena. Artificial neural networks (ANN) Artificial Neural Networks (ANNs) are inspired by the neural neurons of the human brain, consisting of three main layers: input, hidden, and output. ANNs typically use gradient descent optimization to continuously update weights and biases via the backpropagation technique. This process aims to minimize the error between the predicted output and observed data (Nourani, Mogaddam and Nadiri, 2008) . The input (x) and output (Y) are related as per Eq. (6). Furthermore, the sigmoid function was used as the activation function for the network’s nodes, as shown in Eq. (7). Figure 4 illustrates the network’s architecture and the overall flowchart of the ANN process.6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y=f\left({W}_{1}{X}_{1}+{W}_{2}{X}_{2}+\cdots +{W}_{n}{X}_{n}+b\right)$$\end{document}7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$f(y)=\frac{1}{1+{e}^{-y}}$$\end{document} Metaheuristic algorithms employed Metaheuristic algorithms are a powerful class of optimization techniques that are crucial for solving complex problems that are either impossible or extremely challenging to handle with conventional optimization techniques. These algorithms are often inspired by natural phenomena such as biological evolution, animal social behavior, or physical processes. A powerful hybrid model that can significantly improve prediction performance is created by combining metaheuristic optimization algorithms with artificial neural networks. ANNs are excellent prediction tools due to their natural ability to extract complex and non-linear patterns from data. However, the performance of an ANN depends critically on the correctness of its internal parameters, such as weights and biases. Algorithms like gradient descent are commonly used in the ANN training process (updating weights and biases), but they can converge slowly or become trapped in local optima as a result of poor learning rate selection. As a result, groundwater level prediction will be significantly impacted by this hybrid combination. The model can learn the intricate relationships between different influencing factors (like rainfall, temperature, water abstraction, etc.) and groundwater levels with much greater accuracy if the ANN’s parameters are precisely optimized using metaheuristic algorithms. As a result, groundwater level change estimates become more accurate, prediction accuracy rises, and model error decreases. In the end, more precise forecasts offer a strong basis for improved choices in agricultural planning, water resource management, and tackling climate change issues. The integrated architecture of the ANN and metaheuristic algorithms is illustrated in Fig. 6.Fig. 6Integrated Architecture of ANN and Metaheuristic Algorithms. The map was generated by the authors using EdrawMax Ultimate version 15.1.4. (https://edrawmax.wondershare.com/) and does not require any permissions. Integrated Architecture of ANN and Metaheuristic Algorithms. The map was generated by the authors using EdrawMax Ultimate version 15.1.4. (https://edrawmax.wondershare.com/) and does not require any permissions. Honey badger optimization algorithm (HBO) The Honey Badger Optimization (HBO) algorithm was developed by Hashim et al.44, inspired by the social behavior of honey badgers. To estimate the location of prey, honey badgers use their scent-tracking abilities. They also utilize the honeyguide bird as a guide, ultimately leading them directly to beehives44. Like many metaheuristic methods, the HBO algorithm consists of two phases: exploration and exploitation. The population of candidate solutions (X) is obtained according to Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X=\left[\begin{array}{ccccc}{X}_{11}& {X}_{12}& {X}_{13}& \dots & {X}_{1d}\\ {X}_{21}& {X}_{22}& {X}_{23}& \dots & {X}_{2d}\\ \dots & \dots & \dots & \dots & \dots \\ {X}_{n1}& {X}_{n2}& {X}_{n3}& \dots & {X}_{nd}\end{array}\right]$$\end{document} The algorithm’s initial phase begins with initialization. The relative position of honey badgers is obtained from Eq. (9) as follows:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{j}={\mathrm{LB}}_{i}+{r}_{1}\left({\mathrm{UB}}_{i}-{\mathrm{LB}}_{i}\right),{r}_{1}\in [0,1]$$\end{document} In this relationship, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{UB}}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{LB}}_{i}$$\end{document} represent the upper and lower bounds in the search space, respectively, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{j}$$\end{document}​ denotes the position of the jth honey badger. This expression describes a candidate solution within a population of size n. The intensity of the honey badgers’ activity depends on the prey concentration and its separation distance from the jth honey badger. Therefore, their fast and slow movements are governed by the intensity of the emitted scent, which follows the Inverse Square Law (ISL) and is obtained from Eqs. (10) to (12).10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{j}=\frac{{r}_{2}S}{4\pi {d}_{i}^{2}},{r}_{2}\in [0,1]$$\end{document}11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S={\left({X}_{j}-{X}_{j+1}\right)}^{2}$$\end{document}12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${d}_{j}={X}_{prey }-{X}_{j}$$\end{document} In these relationships, S is an index of the concentration or intensity of a source (prey location), and dj​ is the measured distance between the prey and the jth honey badger. The third step is to update the density factor (α), which controls the time-varying randomness from exploration to exploitation. The reduction process of this update (with a random nature) in each iteration is performed according to Eq. (13):13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha =C\times \mathrm{exp}\left(\frac{-\mathrm{iter}}{{\mathrm{Max}}_{\mathrm{iter}}}\right)$$\end{document} In this relationship, C is a constant greater than one (default is 2), Maxiter represents the maximum number of iterations, and iter is the current iteration. The next step involves taking necessary measures to escape local optima. To this end, a flag (F) is used to change the search direction, providing favorable conditions for a thorough scan of the search field. The subsequent phase is updating the position of the search agents (badgers). There are two common methods for updating the search agent positions (Xnew), which are derived from Eqs. (14) to (11):14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{c}{X}_{new }={X}_{prey }+F.\beta .I.{X}_{prey }+F.{r}_{3}.\alpha .{d}_{j}\\ \left|\mathrm{cos}\left(2\pi {r}_{4}\right)\left[1-\mathrm{cos}\left(2\pi {r}_{5}\right)\right]\right|\end{array}$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{new }={X}_{prey }+\alpha {.d}_{j}.{r}_{7}.F.{r}_{7}\in [0,1]$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$=\left\{\begin{array}{c}1 if {r}_{6}\le 1/2\\ -1 otherwise \end{array}{r}_{6}\in [\mathrm{0,1}]\right.$$\end{document} In these relationships, Xbest​ represents the best global prey position, F denotes the honey badger’s ability to obtain food (default value is 6), and r3​,r4​, and r5​ are random numbers generated within the range of 0 and 1, respectively. Xnew​ and Xprey​ refer to the new and previous prey positions, and dj​ signifies the search distance (the distance between the honey badger and the prey). Pelican optimization algorithm (POA) The POA was developed by Trojovský and Dehghani, inspired by the social behavior of pelican flocks45. A population-based algorithm, the Pelican Optimization Algorithm (POA) considers pelicans to be its members. Every member of the population in population-based algorithms is a potential solution. Based on its location in the search space, each member suggests values for the variables involved in the optimization problem46 . Equation (17) is first used to randomly initialize population members based on the lower and upper bounds of the problem.17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i,j}={l}_{j}+\mathrm{rand}\cdot \left({u}_{j}-{l}_{j}\right),i=\mathrm{1,2},\dots ,N,j=\mathrm{1,2},\dots ,m$$\end{document} Equation (18) illustrates how the population members (pelicans) in the proposed POA are identified using a matrix known as the population matrix. This matrix’s columns show the suggested values for the problem’s variables, while each row represents a potential solution.18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$={\left[\begin{array}{c}{X}_{1}\\ \vdots \\ {X}_{i}\\ \vdots \\ {X}_{N}\end{array}\right]}_{N\times m}={\left[\begin{array}{ccccc}{x}_{\mathrm{1,1}}& \cdots & {x}_{1,j}& \cdots & {x}_{1,m}\\ \vdots & \ddots & \vdots & & \vdots \\ {x}_{i,1}& \cdots & {x}_{i,j}& \cdots & {x}_{i,m}\\ \vdots & & \vdots & \ddots & \vdots \\ {x}_{N,1}& \cdots & {x}_{N,j}& \cdots & {x}_{N,m}\end{array}\right]}_{N\times m}$$\end{document} In this context, X represents the pelican population matrix, and Xi​ is the ith pelican. Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i,j}$$\end{document}​ is the value of the jth variable specified by the ith candidate solution. N is the number of population members, m is the number of problem variables, rand is a random number in the range [0, 1], lj​ is the lower bound of the jth variable, and uj​ is the upper bound of the jth problem variable. In the proposed POA, each population member is a pelican, which serves as a candidate solution for the problem at hand. The columns of this matrix display the proposed values for the problem variables, as per Eq. (19).19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X={\left[\begin{array}{c}{X}_{1}\\ \vdots \\ {X}_{i}\\ \vdots \\ {X}_{N}\end{array}\right]}_{N\times m}={\left[\begin{array}{ccccc}{x}_{\mathrm{1,1}}& \cdots & {x}_{1,j}& \cdots & {x}_{1,m}\\ \vdots & \ddots & \vdots & & \vdots \\ {x}_{i,1}& \cdots & {x}_{i,j}& \cdots & {x}_{i,m}\\ \vdots & & \vdots & \ddots & \vdots \\ {x}_{N,1}& \cdots & {x}_{N,j}& \cdots & {x}_{N,m}\end{array}\right]}_{N\times m}$$\end{document} As a result, each potential solution can be used to assess the problem’s objective function. Equation (20) uses a vector known as the objective function vector to calculate the obtained values for the objective function:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F={\left[\begin{array}{c}{F}_{1}\\ \vdots \\ {F}_{i}\\ \vdots \\ {F}_{N}\end{array}\right]}_{N\times 1}={\left[\begin{array}{c}F\left({X}_{1}\right)\\ \vdots \\ F\left({X}_{i}\right)\\ \vdots \\ F\left({X}_{N}\right)\end{array}\right]}_{N\times 1}$$\end{document}where F is the objective function vector and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{i}$$\end{document}​ is the objective function value for the ith candidate solution. The pelican’s strategy for moving towards the prey’s location is mathematically simulated in Eq. (21).21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i,j}^{{p}_{1}}=\left\{\begin{array}{cc}{x}_{i,j}+\mathrm{rand}\cdot \left({p}_{j}-I\cdot {x}_{i,j}\right),& {F}_{p}<{F}_{i};\\ {x}_{i,j}+\mathrm{rand}\cdot \left({x}_{i,j}-{p}_{j}\right),& \text{ else} ,\end{array}\right.$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i,j}$$\end{document}​ is the new position of the ith pelican in the jth dimension based on Phase 1. I is a random integer that can be either 1 or 2. Pj​ is the prey’s position in the jth dimension, and FP​ is its objective function value. The parameter I is a randomly generated number equal to either 1 or 2. The new position for a pelican is obtained according to Eq. (22) as follows:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{i}=\left\{\begin{array}{cc}{X}_{i}^{{P}_{1}},& {F}_{i}^{{P}_{1}}<{F}_{i}\\ {X}_{i},& \text{ else}\end{array}\right.$$\end{document}23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i,j}^{{P}_{2}}={x}_{i,j}+R\cdot \left(1-\frac{t}{T}\right)\cdot (2\cdot \, rand \, -1)\cdot {x}_{i,j}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{i}^{{P}_{1}}$$\end{document}​ is the new position of the ith pelican, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{i}^{{P}_{1}}$$\end{document}​​ is its objective function value based on Phase 1. Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i,j}^{{P}_{2}}$$\end{document}​ is the new position of the ith pelican in the jth dimension based on the second phase. R is a constant equal to 0.2. Also, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R\cdot \left(1-\frac{t}{T}\right)$$\end{document}​ is the neighborhood radius, while t is the iteration counter and T is the maximum number of iterations. Effective updating is also used to accept or reject the pelican’s new position, as modeled in Eq. (24).24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{i}=\left\{\begin{array}{cc}{X}_{i}^{{P}_{2}},& {F}_{i}^{{P}_{2}}<{F}_{i};\\ {X}_{i},& \text{ else} ,\end{array}\right.$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X}_{i}^{{P}_{2}}$$\end{document}​ is the new position of the ith pelican, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F}_{i}^{{P}_{2}}$$\end{document}​​ is its objective function value based on Phase 2. Figure 7 illustrates the overall flowchart of the HBO and POA metaheuristic algorithms.Fig. 7Operational Flowchart of HBO and POA Algorithms. The map was generated by the authors using EdrawMax Ultimate version 15.1.4. (https://edrawmax.wondershare.com/) and does not require any permissions. Operational Flowchart of HBO and POA Algorithms. The map was generated by the authors using EdrawMax Ultimate version 15.1.4. (https://edrawmax.wondershare.com/) and does not require any permissions. Evaluation metrics for model performance To accurately assess a model’s precision, it’s essential to use various evaluation metrics to examine the results. For this purpose, we utilized Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Coefficient of Determination (R2), Nash–Sutcliffe Efficiency Coefficient (NSE), and Mean Absolute Percentage Error (MAPE). A summary of the relationships’ limitations is provided in Table 4.Table 4Evaluation metrics for model performance.Metric nameMathematical formulaRange of definitionOptimum valueMean absolute error (MAE)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAE=\frac{1}{n}\sum_{i=1}^{n} \left|{A}_{i}-{P}_{i}\right|$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left(0