Data-Driven LSTM Architectures for Reservoir Inflow Forecasting

Accurate forecasting of reservoir inflow is crucial for managing water resources, maintaining a balance between water supply and demand, preventing floods, supporting hydropower production, and planning irrigation. India, ranking third globally, with more than 5,000 dams, faces challenges in reservoir operations due to hydrological variability caused by the monsoon. While ensuring demand, supply, and flood security requires high water levels, the reservoir also needs to maintain a certain amount of free storage to accommodate high inflows. While Long-Short Term Memory (LSTM) models have been widely used for inflow forecasting, traditional LSTM models often limit their ability to capture sudden hydrological extremes and accurately represent peak timings. Therefore, a comparative evaluation of various advanced LSTM variants is necessary to identify architectures that are more reliable for modelling nonlinear inflow dynamics. Our study introduces a specialised type of recurrent neural network, specifically the LSTM framework, for forecasting daily reservoir inflow. Our methodology uses a structured feature engineering strategy that integrates hydrometeorological forcings, hydrological state variables, and outputs from the CaMa-Flood hydrodynamics model. A permutation-based feature importance analysis, in terms of the increase in mean absolute error, highlights that antecedent precipitation and lagged upstream reservoir outflow are the main influencing factors for the inflow forecast within a multivariate sequence-to-one LSTM framework. Overall, this framework provides a strong, scalable, and practical solution for inflow forecasting. By supporting timely operational decisions for water release, flood preparedness and storage optimisation, the framework serves as an effective tool for managing reservoirs.