LSTM (Long Short-Term Memory)-based Hydrograph Modeling for Pandanduri Reservoir, Indonesia

Abstract

Dams are vital infrastructure that supports irrigation, raw water supply, flood control, hydroelectric power, and tourism. Reservoirs of the dam store the water during the rainy season and release it in the dry season to reduce the flood risk. Continuous monitoring is essential to ensure their optimal function and safe operation. However, in many developing regions, the availability and quality of discharge data remain inadequate due to poor data governance, limited resources, and insufficient digital infrastructure. These challenges highlight the need for new approaches capable of improving discharge prediction under limited data conditions. To address this issue, Machine Learning (ML) has gained popularity in hydrological research. Long Short-Term Memory (LSTM) is one of the ML methods that has proven effective for analyzing historical data patterns for prediction. This study uses an LSTM-based hydrograph model in the Pandanduri Reservoir Watershed, Indonesia. Compared to conventional statistical or conceptual models, LSTM offers the advantage of capturing nonlinear dynamics in long time-series datasets. The model utilized open data, including satellite rainfall data, meteorological data, and streamflow data records, to capture the complex relationship between rainfall and streamflow.  The generated model was validated with the recorded streamflow data. The model evaluation produces an NSE of 0.44 and an RMSE of 0.74. These values indicate that the model is capable of reproducing key rainfall–runoff dynamics under data-limited conditions. Further research is essential to improve the ability of LSTM to capture extreme events and improve its generalization across hydrological conditions.