Forecasting daily rainfall in Coastal Ghana with machine learning: an evaluation analysis of gated residual learning framework

Accurate rainfall prediction is critical for climate adaptation, agriculture, water resource management, and other arenas, especially in tropical monsoon regions like Coastal Ghana. Rainfall exhibits strong seasonality, non-linearity, and high interannual variability in these areas. Traditional statistical models such as ARIMA struggle to capture these complex dynamics, especially during extreme rainfall events. Machine learning (ML) approaches offer flexible, non-parametric alternatives that can capture non-linear relationships. However, hybrid ML approaches that incorporate gated mechanisms for noise filtering and residual correction remain largely unexplored, especially in West African rainfall forecasting. This study proposed a gated residual learning (GRL) framework that integrates sequence-based models [long-short-term memory (LSTM) and Temporal Convolutional Network (TCNs)] and non-sequence learners [Support Vector Regression (SVR), Random Forest, XGBoost, and Gradient Boosting] with a secondary residual learner to correct systematic prediction errors and improve daily rainfall prediction accuracy. Using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R 2 metrics, the performance of the models was evaluated across the full test dataset, specifically during extreme rainfall events (≥90th percentile). Results from the study show that integration of gated residual learning (GRL) significantly enhances the performance of baseline models. Overall, Gated-SVR achieved the best RMSE and the highest R 2 . Gated-XGBoost attained the best MAE. Integration of GRL improves alignment with rainfall events and mitigates systematic bias. This underscores the framework's capacity to refine both classical and deep learning models.