From data to decisions: Leveraging ML for improved river discharge forecasting in Bangladesh

Abstract

River discharge forecasting stands at the forefront of environmental management, contributing significantly to sustainable development through its impact on flood prevention, water resource management, ecological conservation, and energy production. This study forecasted the annual river discharge forecasting in the Nilphamari district of Bangladesh, employing random forest (RF), support vector machine (SVM), and gradient boosting machine (GBM) techniques. Historical river discharge data spanning from 1990 to 2020, obtained from eight surface water stations, forms the basis of the analysis. The forecast was performed from 2021 to 2030. 11 statistical parameters were considered for performance evaluation. Additionally, four evaluation plots, comprising a quantile–quantile plot (QQ plot), a residual plot, a Bland Altman plot, and Theil’s U statistic, were employed for a detailed understanding of model accuracy. Results demonstrate that the random forest regression technique exhibited superior accuracy compared to SVM and GBM in training and testing stages. Notably, the coefficient of determination reached 97 % during the testing phase, emphasizing the robustness of this model. While Mean Absolute Error is lower (1085.071 cubic meter per second), in training, the model captures relative changes (Mean Absolute Percentage Error = 0.154) better during prediction. Willmott’s Index in training (0.77) and testing (0.55) suggest the model memorizes training data well and outperforms the other models in testing stage. The findings underscore the efficacy of RF regression as a superior alternative for short-term discharge forecasting, offering valuable insights for integrated water resources management, particularly in flood warning systems and the expansion of irrigation initiatives.