A hybrid deep learning framework for improving short-term precipitation forecasts

Abstract

Precipitation forecasting is essential for water-resources, flood-risk, and agricultural planning, yet the complexity and uncertainty of the system limit predictive accuracy. This study introduces a hybrid deep learning (DL) model for daily precipitation forecasting, using a gated recurrent unit to capture temporal dependencies and a multilayer perceptron to account for nonlinear effects. Hyperparameter tuning was performed using a Bayesian optimization algorithm during data preprocessing, architecture design, and model training. The framework was evaluated using daily precipitation data from the McCloud Reservoir drainage basin. 68% decrease in mean squared error was obtained compared with the raw forecast baseline; improvements were also observed in Nash–Sutcliffe and Kling–Gupta efficiency indices. Predictive skill was increased for most rainfall intensities, and threat-score and bias metrics show clear gains, even for extreme events. Results show that the proposed hybrid DL can improve rainfall prediction and support reservoir operations, flood preparedness, and climate-resilient water management.