The study demonstrates a multi-model ensemble method for river discharge forecasting over flood-prone major rivers of India (Brahmaputra, Ganga, and Kosi). Deep learning-based models (RNN, GRU, LSTM, BiLSTM) are applied over these diverse river systems of India. These models are trained using rainfall, soil moisture, and model-simulated river discharge with 3-day, 5-day, and 7-day moving training windows (T-3, T-5, and T-7) to generate 1 to 7 day (F-1 \(\ldots \) F-7) discharge forecasts. To enhance forecast performance, an ensemble model approach is proposed in this study. A global Ridge model is used, which takes deep learning model outputs with statistical features from the data to generate multi-day river discharge forecasts. Performance evaluation is carried out for individual deep learning models and the ensemble global Ridge model. Individual models performed well for Day-1 forecasts across all sliding temporal training windows with statistical measures such as Nash–Sutcliffe efficiency performing well (NSE \(\approx \) 0.9–0.97) for all river systems. However, performance degrades with increase in lead time in all deep learning models. The ensemble model improves overall performance compared to individual models across all training windows, especially for longer lead times. The results for the Day-1 ensemble model forecast reflect higher performance over the Brahmaputra (NSE = 0.974), Ganga (NSE = 0.966), and Kosi (NSE = 0.978) rivers. Substantial improvement is observed for Day-7 with NSE values of 0.711 (Brahmaputra), 0.845 (Ganga), and 0.830 (Kosi) for the T-7 training window. These results highlight that combining different deep learning models with varying architectures by a global Ridge ensemble model yields robust short-to-medium range discharge forecasts in data-sparse river basins, providing a promising, computationally efficient tool for the development of an operational flood early-warning system.
