Comparative Evaluation of Transformer and Machine Learning Algorithms for Predicting Downstream River Stages During Spillway Release Events

Abstract

Predicting downstream river stages during extreme reservoir operations remains challenging due to complex interactions between dam operation and river response. Although machine learning has shown potential for hydrological applications, there remains a gap in the comparative evaluation of Transformer-based and conventional architectures under high-spillway-discharge conditions. This study proposes a Transformer-based regression model to estimate daily maximum stages at two stations downstream of the Urrá I Hydroelectric Power Station (Colombia), using reservoir inflow, turbine and spillway discharges, and antecedent river stages. The Transformer was implemented in a point-wise configuration (T=1). Using 117 daily observations, chronological training–validation splitting was evaluated using five training proportions (50–90%), with 10 random seeds per configuration, and comparisons were made with LSTM, Random Forest, and XGBoost. The Transformer achieved the strongest performance at 50% and 60% training proportions, while the other algorithms became more competitive at larger proportions. Overall, the Transformer demonstrated competitive predictive capability, particularly under limited training data availability.