Water demand forecasting based on multi-rainfall gauging stations using stand-alone soft computing techniques with improved novel hybrid paradigms

Water demand forecasting is crucial for effective water resource planning and management. This study presents a reliable and cost-effective approach to model and predict water demand using rainfall measurements from various stations. Four stand-alone methods were employed: Support Vector Regression (SVR), Gaussian Process Regression (GPR), Least Square Boost (LSBOOST), and Stepwise Linear Regression (SWLR). To enhance prediction performance, novel hybrid models were developed, combining SWLR with SVR, GPR, and LSBOOST. The effectiveness of both the stand-alone and hybrid methods was assessed using four statistical metrics: Nash-Sutcliffe efficiency (NSE), Root Mean Square Error (RMSE), Pearson Correlation Coefficient (PCC), and Mean Absolute Percentage Error (MAPE). Results showed that none of the stand-alone techniques could effectively predict the water demand due to its complex nature. In contrast, the hybrid techniques, particularly SWLR-GPR and SWLRLSBOOST, demonstrated robust performance, achieving a minimum NSE of 0.95 during both calibration and validation. Graphical analyses, including time series plots and a 2-dimensional Taylor diagram (2D-TD), illustrated these models' ability to capture daily demand fluctuations. The hybrid models significantly outperformed the stand-alone techniques, improving prediction accuracy by over 81% to 88% in calibration and validation phases, respectively, despite relying solely on rainfall as input. This study underscores the potential of hybrid modeling approaches in enhancing water demand forecasting accuracy.