TeaNet: An Enhanced Attention Network for Climate-Resilient River Discharge Forecasting Under CMIP6 SSP585 Projections

The accurate prediction of river discharge is essential in water resource management, particularly under variability due to climate change. Traditional hydrological models commonly struggle to capture the complex, nonlinear relationships between climate variables and river discharge, leading to uncertainties in long-term projections. To mitigate these challenges, this research integrates machine learning (ML) and deep learning (DL) techniques to predict discharge in the Subernarekha River Basin (India) under future climate scenarios. Global climate models (GCMs) from the Coupled Model Intercomparison Project 6 (CMIP6) are assessed for their ability to reproduce historical discharge trends. The selected CNRM-M6-1 model is bias-corrected and downscaled before being used to simulate future discharge patterns under SSP585 (a high-emission scenario). Various AI-driven models, such as a temporal convolutional network (TCN), a gated recurrent unit (GRU), a support vector regressor (SVR), and a novel DL network named the Temporal Enhanced Attention Network (TeaNet), are implemented by integrating the maximum and minimum daily temperatures and precipitation as key input parameters. The performance of the models is evaluated using the mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and coefficient of determination (R2). Among the evaluated models, TeaNet demonstrates the best performance, with the lowest error rates (RMSE: 2.34–3.04; MAE: 1.13–1.52 during training) and highest R2 (0.87–0.95), outperforming the TCN (R2: 0.79–0.88), GRU (R2: 0.75–0.84), SVR (R2: 0.68–0.80), and RF (R2: 0.72–0.82) by 8–15% in accuracy across four gauge stations. The efficacy of the proposed model lies in its enhanced attention mechanism, which successfully identifies temporal relationships in hydrological information. In determining the most relevant predictors of river discharge, the feature importance is analyzed using the proposed TeaNet model. The findings of this research strengthen the role of DL architectures in improving long-term discharge prediction, providing valuable knowledge for climate adaptation and strategic planning in the Subernarekha region.