Water Level Forecasting Using a Transformer-Based Deep Learning Algorithm for the Baitarani Basin in India

Abstract Water level forecasting plays a pivotal role in flood risk mitigation, reservoir operation, and water resource management, particularly in hydrologically sensitive regions like the Baitarani River basin in eastern India. This study investigates the performance of three deep learning models—long short-term memory (LSTM), convolutional neural network (CNN), and Temporal Fusion Transformer (TFT)—for short-term water level prediction, with a focus on lead times ranging from 1 to 24 h. The novelty of this work lies in assessing the recently developed TFT model’s ability to integrate attention-based temporal dynamics for hydrological forecasting, an area rarely explored in Indian river basins. The Baitarani basin, characterized by seasonal monsoonal floods and complex hydrological dynamics, requires accurate and timely forecasts to support early warning systems and emergency response planning. Historical hydrometeorological data were preprocessed and used to train, optimize, and validate the models. The results in terms of evaluation matrices reveal that while all three models demonstrate satisfactory predictive performance for short-term forecasts, the TFT significantly outperforms both LSTM and CNN models, particularly at longer lead times. At a 24-h forecast horizon, the TFT model achieved a root-mean-square error (RMSE) of 0.446 m, a mean absolute error (MAE) of 0.312 m, and an Nash–Sutcliffe efficiency (NSE) of 0.803 during validation—superior to the LSTM and CNN models, which recorded RMSE values of 0.5034 m and 0.5029 m, respectively. The superior performance of the TFT model can be attributed to its attention mechanism, temporal gating components, and ability to handle multivariate inputs efficiently. These characteristics enable the model to capture both short-term fluctuations and long-term dependencies in hydrological time-series data more effectively than traditional deep learning architectures. However, its generalization may be constrained by data resolution and regional hydrological conditions, suggesting that further validation across diverse basins is necessary.