Short‐Term Reservoir Inflow Forecasting in an Arid Flash‐Flood System

Résumé fourni par la source

ABSTRACT Short‐term reservoir inflow forecasting is essential for flood‐risk mitigation and operational decision‐making in arid regions, where runoff is highly intermittent and dominated by short‐lived flash floods. This study compares persistence, multivariate linear regression and a Long Short‐Term Memory (LSTM) network for forecasting inflow to Najran Dam, Saudi Arabia, using hourly rainfall and reservoir inflow records for 2016–2020. Models were trained and evaluated using a strict chronological framework at lead times of 1, 2, 4 and 6 h. Forecast performance was assessed using root mean square error, mean absolute error and Nash–Sutcliffe efficiency. Forecast skill decreased rapidly with increasing lead time for all models. Linear regression consistently outperformed persistence beyond the 1‐h horizon, demonstrating the value of combining rainfall forcing with recent inflow observations. In contrast, the LSTM did not provide systematic improvement over the simpler and more interpretable linear‐regression model under the available 5‐year record. At the longer tested lead times, both approaches converged toward low skill, reflecting the combined influence of data sparsity, extreme inflow skewness, limited rainfall information and rapid catchment response. Useful forecast skill under the tested dataset, predictors and model configurations was concentrated within approximately 1–4 h. These findings emphasise the importance of benchmarking complex models against simple and interpretable alternatives when developing operational reservoir‐inflow forecasting systems from short hydrological records.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.