Advanced Framework for Predicting Rainfall-Runoff: Comparative Evaluation of AI Models for Enhanced Forecasting Accuracy

The accurate prediction of river discharge, particularly inflow to reservoir dams, is necessary for effective water resources management and planning. This research focuses on monthly rainfall-runoff modelling for the Boukan dam, a sub-basin of Urmia lake, utilizing five distinct AI techniques: artificial neural networks (ANN), gene expression programming (GEP), least squares support vector machine (LSSVM), multivariate adaptive regression spline (MARS), and random forest (RF). Additionally, an ensemble machine learning technique was applied to evaluate potential improvements in runoff estimation. For this purpose, monthly rainfall and discharge data from 1991 to 2022 were collected from three river and rain gauging stations located on separate rivers feeding into the Boukan Dam reservoir. Various input combinations were decided using the partial auto-correlation function, cross correlation function and stepwise regression. The efficiency metrics (R2, RMSE, NSE, LMI) for the preeminent AI models at the Pol-Aienan, Safa-khaneh, and Sonateh stations were determined as (0.905, 7.423 m3/day, 0.900, 0.737), (0.822, 4.724 m3/day, 0.817, 0.570), and (0.812, 5.290 m3/day, 0.810, 0.633), respectively. The SHapely Additive exPlanation (SHAP) algorithm revealed that the current month’s precipitation and the previous month’s discharge were the most significant variables influencing current month runoff across all stations. These findings affirm the AI methods’ proficiency in accurately capturing the runoff at three river stations.