Enhancing Rainfall Forecasting in Tunisia: Application of a Hybrid Deep Learning Approach

Accurate rainfall data are essential for hydrological forecasting and climate modeling. However, many developing regions, including Tunisia, struggle with significant data gaps in rainfall measurements, particularly from gauge stations. These missing data impair climate model validation and reduce forecasting accuracy across both spatial and temporal dimensions. To overcome these limitations, we conduct a comprehensive evaluation of novel deep learning (DL) architectures designed for imputing missing rainfall gauge data and generating monthly rainfall forecasts. Our framework systematically compares multiple DL approaches: Long Short-Term Memory (LSTM), a hybrid Bidirectional LSTM with a Transformer attention mechanism (BiLSTM-Transformer), and a pure Transformer model. Subsequently, we employ Principal Component Analysis (PCA), K-Means clustering, and quantile techniques to further refine DL model outputs. The processed data are then analyzed using Light Gradient Boosting Machine (LightGBM) to produce final results. Our rigorous evaluation across 47 Tunisian gauges covering 1983–2012 (70% training, 30% testing) demonstrates that the BiLSTM-Transformer hybrid delivers superior performance, achieving an 18.4% reduction in root mean squared errors (RMSE) compared to conventional interpolation methods (14.2 mm versus 17.4 mm monthly error) and improving R2 values by 0.15–0.23 across all test stations. The model shows particular strength in capturing Mediterranean rainfall patterns, correctly predicting 83% of extreme rainfall events (greater than 95th percentile). Furthermore, spatial graph networks boost performance at data-sparse stations by 12.7% through explicit modeling of topographic influences.