Novel deep learning framework for rainfall forecasting integrating generative adversarial and spatiotemporal graph neural networks

Global warming and climate change cause intense extreme precipitation, demanding advanced rainfall prediction techniques for improved water resource planning and management. This study developed two deep-learning algorithms, Spatio-Temporal Graph Neural Networks (ST-GNNs) and Generative Adversarial Graph Neural Networks (GA-GNNs), and evaluates their performance on the entire meteorological data of Thailand in the years 1993 to 2022. The findings indicate that ST-GNNs outperform GA-GNNs in terms of accuracy: ST-GNN has a high coefficient of determination (R2 = 0.9580), a low root mean square error (RMSE = 0.1560), and mean absolute error (MAE = 0.3950), which proves its ability to capture complex spatiotemporal rainfall patterns. In contrast, GA-GNN has poor performance measures (R2 = 0.5800, RMSE = 0.8400, MAE = 1.5770) that confirm the fact that it has little representational capabilities over the intricate rainfall dynamics. SHAP analysis has shown that wind speed, relative humidity, and minimum temperature have the most significant impact on ST-GNN predictions, and wind speed is the most positive factor (SHAP value = 0.634). Moreover, the paper records that STGNNs are superior when it comes to rainfall category classification, with a weighted F1 score of 1.00 and an accuracy of 0.992- a difference that is much better than that of GA-GNNs. The results highlight the importance of considering both spatial and temporal dependencies in modeling rainfall in climatically complex environments like Thailand.