COMPARATIVE EVALUATION OF MACHINE LEARNING-BASED MODELS FOR RAINFALL–RUNOFF PREDICTION

Abstract

Accurate rainfall–runoff prediction is important in hydrological modeling. However, watershed processes are nonlinear and dynamic. This study examined the effect of different input configurations on daily streamflow prediction using data-driven methods. A total of 1,826 daily observations covering the period 01/01/2020–09/13/2025 were used. The dataset was split into 1,461 for training and 365 for testing. Three model sets were developed and evaluated using Multiple Linear Regression (MLR), Multilayer Perceptron (MLP), M5 Model Tree (M5 TREE), and Random Tree (RT). Model performance was assessed using RMSE, MAE, MAPE, NSE, R, and R². The results showed that prediction accuracy improved as the input structure became more comprehensive. Model Set 3 produced the best overall results. MLP3 achieved the lowest RMSE (3.582 m³/s) and the highest NSE (0.974) and R² (0.974). M5 TREE3 produced the lowest MAE (2.118 m³/s) and MAPE (4.601%). In contrast, RT models showed weaker and less stable performance. Overall, the results showed that input configuration is as important as algorithm selection in rainfall–runoff modeling. The findings also indicated that nonlinear machine learning models can provide more reliable predictions when suitable lagged variables are included.