Enhanced dam spillage prediction using lagged feedback random forests at Weija Dam, Ghana

Abstract

Accurate prediction of dam spillage is essential for mitigating flood risks to downstream communities and for the optimal management of water resources. Existing machine learning models often struggle to capture the delayed hydrological responses that influence reservoir spillage. This limits their operational usefulness for early warning and reservoir management. This study proposes a lag-enhanced Random Forest forecasting framework, termed Lagged-Feedback Random Forest (LFRF), that systematically incorporates temporal feedback from historical data points to improve dam spillage prediction. The framework was benchmarked against three alternative approaches, Long Short-Term Memory (LSTM) networks, standard Random Forest (RF), and Artificial Neural Networks (ANN), as well as four additional contemporary architectures, XGBoost, LightGBM, CatBoost, and a Temporal Convolutional Network (TCN), using the Weija Dam's 15-year daily historical record and a 40-day lag structure. Models were evaluated using RMSE, MAE, and R² (in m³/s, consistent with the observed spillage discharge), alongside the dimensionless Correlation and T-test statistics. LFRF outperformed all other models with an R² of 0.94 and a Correlation of 0.97, with a T-test statistic of 0.078 indicating minimal systematic deviation from observed values; LFRF and LSTM achieved comparable, markedly lower prediction errors than RF and ANN. Standard RF, XGBoost, LightGBM, CatBoost, and ANN were evaluated as naive baselines using only current-day readings, without either the lagged features given to LFRF or the sequence-native architecture used by LSTM and a Temporal Convolutional Network; all five naive models performed similarly poorly regardless of algorithm sophistication. For the Weija Dam dataset and the configurations evaluated, access to lagged temporal information appeared to contribute more strongly to predictive performance than model architecture alone. These findings highlight the value of lagged features in Random Forest ensembles for dam spillage prediction and identify several lag “sweet spots” — 7-8 days, 15-17 days, 25-26 days, and approximately 30-40 days — at which LFRF provided consistently strong predictive performance that could support real-time operational decisions and risk mitigation at the dam.