Streamflow forecasting using LSTM and weighted curve number method in Tawi watershed, Western Himalaya

Abstract This study develops a hybrid hydrological forecasting framework that integrates a physically based Weighted Curve Number (WCN) method with a Long Short-Term Memory (LSTM) network to improve streamflow prediction in the Tawi watershed, Western Himalaya. The research aims to enhance the physical interpretability of deep learning models while maintaining strong predictive capability in data-scarce mountainous environments. Daily rainfall, temperature, and discharge data from 2000–2020 were used to train and validate the model at Jammu and Udhampur gauging stations. The WCN-derived runoff estimate was incorporated as an additional input feature within the LSTM architecture to account for land-use and soil-controlled hydrological responses. The hybrid model demonstrated strong performance, achieving R2 values between 0.84 and 0.86 during training and testing. Future projections for 2021–2040 indicate stable seasonal discharge patterns at Jammu, whereas Udhampur shows declining mean flows and increased low-flow sensitivity based on Flow Duration Curve analysis. By combining process-based runoff estimation with deep learning, the proposed WCN–LSTM framework improves predictive robustness and provides a transferable tool for long-term water resource planning, drought risk assessment, and climate adaptation in Himalayan River basins.