Accurate monthly precipitation forecasting holds significant importance for agriculture, meteorological prediction, and environmental protection. While traditional models like Vector Autoregression (VAR) have been widely applied in river flow prediction, their limitations in addressing spatial attributes of meteorological data remain notable. To address this gap, this study proposes a novel ”decomposition-reconstruction-prediction-integration” framework based on Spatio-Temporal Graph Neural Networks (STGNN), which inherently excels in processing multi-site data. First, the Time-Varying Filter-based Empirical Mode Decomposition (TVF-EMD) is employed to decompose raw precipitation sequences into multiple components. Subsequently, frequency dispersion metrics evaluate sequence volatility, with components exceeding the entropy threshold identified as high-frequency signals. These aggregated high-frequency components undergo secondary decomposition through Variational Mode Decomposition (VMD) to generate refined sub-components. The reconstructed components, formed by integrating these sub-components with residual elements, are then fed into an enhanced STGNN model incorporating temporal attention mechanisms, spatial attention layers, and residual optimization modules. Final precipitation forecasts are obtained by synthesizing predictions from all components. Applied to monthly precipitation data spanning January 1979 to August 2023 across Guangdong Province monitoring stations, this model demonstrates superior reliability and accuracy compared to benchmark methods. The proposed framework effectively captures spatiotemporal dependencies while addressing volatility heterogeneity in precipitation patterns, offering a robust solution for regional hydrological forecasting.

Precipitation forecast,graph neural network,Series decomposition,Time series