One-month-ahead river discharge forecasting using machine learning and limited hydrometeorological data: the Karatoya-Atrai River basin, Bangladesh

Abstract

Bangladesh, a riverine country, relies heavily on its intricate network of rivers and water bodies. With over 700 rivers flowing through its landscape, water resources are integral to the nation’s agriculture, transportation, and overall way of life. Effective management of these resources is vital to mitigating the risks of floods, droughts, and other hydrological hazards. This study specifically investigates the capacity of various Machine Learning (ML) models to forecast monthly discharge of the Karatoya-Atrai River 1 month ahead, using only rainfall and temperature, together with antecedent (lagged) discharge. Through a comprehensive comparison of models involving Support Vector Regression (SVR), Random Forest (RF), Linear Regression (LR), Polynomial Regression (PR), Ridge Regression (RR), and Extreme Gradient Boosting (XGBoost), the study underscores that the RF model offers superior discharge forecasting accuracy. Using a strictly chronological train-test split with blocked time-series cross-validation, the RF gives the best one-month-ahead forecasts (NSE 0.86, RMSE 69.71 m 3 s −1 , MAE 43.97 m 3 s −1 , MAPE 40.26%, KGE 0.89, NRMSE_range 0.11), outperforming persistence (NSE 0.51) and seasonal-climatology (NSE 0.80) baselines. An ablation shows that Khosla-derived runoff, which correlates with rainfall at r = 0.97, adds essentially no independent skill, whereas antecedent-discharge memory contributes most. Regime-specific evaluation shows good seasonal skill but reduced accuracy for high and extreme flows. The main value of the framework is as a low-data, one-month-ahead forecasting tool for a data-scarce monsoonal basin, benchmarked against physically meaningful baselines. This offers potential for adoption in monitoring and control systems with limited resources.