The Impact of Rainfall Pattern Dataset Construction on Neural Network Performance for Reservoir Water Level Forecasting

Abstract

Reservoir water level forecasting is a critical component of effective water resources management, supporting flood mitigation, water supply planning, and sustainable reservoir operation, particularly under increasingly variable rainfall conditions. During periods of heavy rainfall, inaccurate or delayed water level prediction may increase flood risk, while during low rainfall seasons, poor forecasting can compromise water storage and operational efficiency. Artificial Neural Networks (ANNs) have been widely adopted for reservoir water level forecasting due to their capability to model nonlinear rainfall–reservoir relationships. However, existing studies largely focus on algorithm selection or architectural enhancement, with limited attention given to how rainfall data representation and dataset construction influence neural network performance. This study addresses this gap by analysing the impact of rainfall pattern dataset construction on ANN performance for reservoir water level forecasting. The primary aim is to evaluate how different rainfall representations affect predictive accuracy when the learning algorithm and training configuration are held constant. Two rainfall pattern datasets were constructed using the same raw rainfall and reservoir water level data from the Timah Tasoh Reservoir, Malaysia. The first dataset represents a compact abstraction of rainfall behaviour using rainfall change indicators derived from day-to-day observations. The second dataset enriches the feature space by incorporating both rainfall change and rainfall intensity categories for each upstream station. In both datasets, the reservoir water level category serves as the prediction target. Prior to model training, redundancy and conflicting data instances were removed to ensure data consistency. A consistent ANN architecture was employed for both datasets and evaluated using 10-fold cross-validation. Model performance was assessed using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). The experimental results demonstrate that the enriched rainfall pattern dataset achieved significantly lower RMSE and MAE values compared to the compact rainfall change dataset, indicating improved learning capability and generalisation performance. Although the enriched dataset required higher computational effort, the improvement in forecasting accuracy was substantial. The findings highlight that dataset construction plays a decisive role in neural-network-based reservoir water level forecasting.