Enhancing Rainfall-Runoff Prediction Accuracy using Artificial Neural Networks: A Case Study of Bida, Nigeria

Abstract

Accurate rainfall-runoff modeling is essential for sustainable water management, especially in flood-prone, data-scarce regions like Nigeria’s Bida Basin. Traditional models struggle with nonlinear hydrological dynamics and limited data availability. This study addresses this gap by developing and evaluating an ANN-based runoff prediction model using limited meteorological and hydrological data. The objective is to improve flow forecasting accuracy and demonstrate the effectiveness of data-driven approaches for climate-resilient water resource planning in under-monitored basins. Daily rainfall, temperature, and runoff data (2010–2023), data were preprocessed, normalized, and partitioned for ANN modeling. A multi-layer ANN was trained using the Adam optimizer and evaluated with RMSE, R², and NSE. The Activation functions (LOGSIG, PURELIN, TANSIG) were tested to assess model accuracy in simulating runoff under nonlinear rainfall-runoff relationships. The ANN model achieved strong runoff prediction performance in the Bida Basin, with R² values of 0.91 (training) and 0.87 (testing), and RMSE of 3.25 and 4.18 m³/s, respectively. PURELIN activation yielded perfect correlation (R = 1.0; RMSE = 0.0), outperforming LOGSIG (R = 0.9995) and TANSIG (R = 0.9547). Seasonal analysis showed higher accuracy in the wet season (R² = 0.89; RMSE = 3.90 m³/s) than in the dry season (R² = 0.77; RMSE = 4.65 m³/s), confirming the model’s robustness across hydrological conditions. ANN models outperform traditional MLR in capturing nonlinear runoff dynamics but risk overfitting without careful tuning, while linear regression excels in simple linear cases, highlighting the need to balance model complexity and generalization based on data and process characteristics.