Delta feature and random forest–enhanced LSTM–attention forecasts with probabilistic postprocessing for rainfall forecasting

Reliable rainfall forecasts are crucial for real-time flood control and managing water resources. Advances in deep learning, high-performance computing, and Internet of Things (IoT)-based synoptic measurements are making rainfall predictions increasingly accurate. The contributions of this work are threefold: (1) leveraging information from neighboring stations with variable time lags to model spatial propagation and regional rainfall dynamics; (2) introducing delta-based features that capture temporal gradients in meteorological variables; and (3) developing a hybrid Attention-Long Short-Term Memory (Attention-LSTM) framework with probabilistic postprocessing for extreme value forecasting. Data from the target station and its seven neighboring stations, along with six meteorological variables, were considered. To capture temporal dependencies, delta and time-lagged features were created, and the most informative inputs were then selected using Principal Component Analysis (PCA), correlation analysis, and Random Forest. The LSTM model was used as a baseline, and several enhanced models were evaluated, including Attention-LSTM, Progressive Graph Convolutional Network-LSTM (PGCN-LSTM), and PGCN-LSTM with attention. The models tended to underestimate extreme rainfall events, likely due to the imbalance in the number of extreme samples. These values were then refined using threshold-based postprocessing with a Weibull distribution. In terms of hyperparameter tuning of the models, regularization and dropout played a key role in preventing overfitting, ensuring robust training of the deep learning framework for short-term rainfall forecasting. The proposed hybrid framework provides satisfactory predictions for general and extreme rainfall (correlation ≈ 0.69), while high-intensity convective events remain partially underestimated.