Metaheuristic-optimized SVR models for daily streamflow forecasting in the Karkheh River Basin, Iran

Estimating river flow at hydrometric stations to determine river discharge in future periods is a significant issue, typically addressed using hydrological time series data from the station itself. In this study, flow discharge data with varying time lags from the Cham Anjir, Kashkan, Pol Zal, and Jologir rivers were utilized under four combined scenarios from 1993 to 2023 for river flow prediction purposes. For modeling, hybrid intelligent models were employed, including Wavelet-Support Vector Regression (WSVR), innovative Gunner Algorithm-Support Vector Regression (AIG-SVR), and Black Widow Optimization-Support Vector Regression (BWO-SVR). Model performance was assessed using the correlation coefficient (R 2 ), root mean square error (RMSE), Normalized Root Mean Square Error (NRMSE), Mean absolute percentage error (MAPE), and Kling–Gupta efficiency (KGE). Furthermore, scatter plots, relative error analysis, and Taylor diagrams were used for results analysis. The results indicated that the combined scenario, incorporating all input data, demonstrated higher accuracy across the examined models. Specifically, the Wavelet-Support Vector Regression (WSVR) model showed superior performance during validation, achieving correlation coefficients R 2 ranging from 0.70 to 0.985, RMSE values from 0.357 to 0.472 m³ /s, MAPE values from 0.171 to 0.236 m³ /s, KGE values from 0.171 to 0.236 m³ /s, NRMSE values from 0.171 to 0.236 m³ /s, and Nash-Sutcliffe efficiency (NSE) values from 0.980 to 0.990. Overall, the findings demonstrate that hybridizing SVR with signal processing techniques like wavelet transform is particularly effective, providing a robust tool for water resource managers in semi-arid regions facing increasing hydrological variability.