ST-GRF: Spatiotemporal graph neural networks for rainfall forecasting

Accurate and real-time rainfall forecasting plays an important role in the weather forecast system and is of great significance for travel planning, engineering planning, and crop management. However, rainfall forecasting has always been considered an open scientific issue, owing to the constraints of geographic topology structure, namely, spatial dependence and temporal dependence. Besides, there are long-term dependencies between climate eigenvalues and the correlation between the atmosphere-related signals at the measuring stations. Traditional approaches usually only focus on temporal dependencies. To capture the spatial and temporal dependence simultaneously, and take into account long-range dependencies and correlations between meteorological signals, we propose a novel neural network-based rainfall forecasting method, the spatiotemporal graph neural networks (ST-GRF) model. In the first step, we innovatively model the generation pattern of rainfall. In the second step, the GRF Module of ST-GRF is used to capture complex topological structures and the dependencies between climate factors. In the third step, the bidirectional temporal dependence capture module of ST-GRF is employed for rainfall forecasting based on spatiotemporal information. Experiments demonstrate that our ST-GRF model can obtain the spatiotemporal correlation from rainfall data and the predictions outperform state-of-art baselines on real-world rainfall datasets. We also discuss and demonstrate the effectiveness of each module in the ablation experiment section. •A framework for multiple data-driven spatiotemporal forecasting is proposed.•A dual-filter graph model is used to establish a rainfall forecasting model.•A dynamic graph generation module Aerodynamic Propagation Fitting can be used in the field of meteorology.•Space-influenced difference mechanism and Bidirectional temporal dependence capture module are proposed.