Enhancing River Discharge Forecasting With Multivariate LSTM

Accurate river discharge forecasting is crucial for water resource management and flood prevention. This study applies Long Short-Term Memory (LSTM) networks to forecast discharge at Neeleshwaram station, Periyar, Kerala, India. Using 18 years of data, exploratory analysis and feature engineering were conducted to select optimal predictive features and determine precipitation-discharge lag time. LSTM models with 1 to 15-day input windows were developed to forecast 5-day discharge. Performance was evaluated using R-squared and RMSE. For 1-day forecasts, a 2-day input window performed best ($\mathrm{R}^{2}=0.82$), highlighting immediate runoff importance. A 6-day window showed best overall performance across all 5 forecast days, balancing short-term responsiveness and longer-term processes. The study demonstrates LSTM networks’ potential in capturing complex hydrological dynamics and provides insights into optimal input window selection, with implications for improving short-term forecasts and water resource management.