Multivariate LSTM-Based Hydrological Modeling for Sustainable Water Resource Management: A Case Study of the Rawa Pening Watershed, Central Java

The sustainable management of water resources requires accurate prediction of hydrological dynam-ics under changing climatic conditions. This study develops a multivariate Long Short-Term Memory (LSTM) model to predict key hydrological variables in the Rawa Pening watershed, Central Java, as part of an effort to support data-driven water management strategies. The model is designed to fore-cast precipitation and surface runoff, which are influenced by meteorological factors including tem-perature, dew point, humidity, radiation, and evaporation. The LSTM framework was selected for its capacity to capture long-term dependencies and nonlinear temporal relationships within multivariate time series. Several combinations of epoch and batch size were tested to determine the optimal model configuration. The best performance for precipitation prediction was obtained with a batch size of 16 and 140 epochs, yielding an RMSE of 24.29 mm and an R² of 0.8333. For runoff prediction, optimal results were achieved with a batch size of 32 and 80 epochs, resulting in an RMSE of 45.59 mm and an R² of 0.8443. The model effectively captures seasonal variability and demonstrates stable predic-tive performance through K-fold cross-validation. Furthermore, a positive correlation between pre-cipitation and temperature, and a negative correlation with radiation, highlights the influence of cli-matic variables on hydrological processes. These findings confirm the potential of LSTM-based pre-dictive modeling as an effective approach to enhance climate-resilient water resource management and can be integrated into decision support systems (DSS) for adaptive watershed planning in tropical lake ecosystems.