IoT-integrated deep learning for forecasting and decision support in reservoir water management under drought conditions

Abstract

This study presents an IoT-enabled forecasting and decision-support framework for proactive reservoir management under drought conditions. Using more than two decades of high-resolution hydrometeorological data, we develop and compare Long Short-Term Memory (LSTM) and extended LSTM (xLSTM) models. The xLSTM integrates exponential gating mechanisms to better capture long-range temporal dependencies. We evaluate predictive performance across multiple forecasting horizons (30, 90, 180, and 365 days) and benchmark the results against a classical statistical model (ARIMA). The xLSTM consistently outperforms baseline models in short-term forecasts but exhibits a decline in accuracy at longer horizons, highlighting the limitations of purely data-driven approaches for extended predictions. To operationalize model outputs, we integrate the forecasts into a real-time decision-support dashboard that aligns predictions with reservoir operation thresholds established in the Catalan Drought Management Plan. This research provides both a methodological contribution to deep learning for hydrological forecasting and a practical framework for data-driven drought preparedness in climate-sensitive regions.