HDA-DGCN: Hierarchical data-driven aggregation network assisted dynamic graph convolutional framework for meteorological prediction

Abstract

Meteorological forecasting, at its core, serves as an indispensable safeguard in our daily lives by accurately predicting atmospheric changes and environmental factors. With the rapid advancement of deep learning technologies in recent years, researchers in this field have begun to widely adopt such methods to achieve higher prediction accuracy. However, current research methods often focus on the analysis of variables at individual meteorological stations within specific regions, neglecting the interactive effects of meteorological observation data across regions and failing to generalize the comprehensive relationships among meteorological variables. Consequently, predicting variables directly without an in-depth understanding is insufficiently comprehensive. To overcome this limitation, we introduce a novel solution: the Hierarchical Data-driven Aggregation Network-assisted Dynamic Graph Convolutional Network (HDA-DGCN). This framework deeply integrates the spatial layout of meteorological stations and the interdependencies of weather variables by considering the multi-level connections and interactions of meteorological stations. It incorporates the Heterogeneous Station Association Mapping Capture (HSAMC) and Dynamic Graph-assisted Graph Convolutional Network (DGDL) modules, enabling in-depth analysis from macro to micro levels. The HSAMC module is data-driven, providing insights into the relationships between stations and the temporal dependencies of weather variables, while the DGDL module generates dynamic graphs through a designed spatiotemporal attribute-based random walk similarity, capturing the spatiotemporal attributes between meteorological variable nodes. Our experimental studies on the station data of Beijing (WD_BJ) and East China (WD_EC) have revealed the significance of the number of stations and key meteorological variables (such as precipitation, temperature, humidity, air pressure, wind, etc.) in optimizing prediction performance. Compared with six well-known baseline models (including DUQ [1], MTGNN [2], CLCSTN [3], etc.) among the nine indicators, the framework we proposed has demonstrated outstanding performance. Specifically, the main indicators, Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), have decreased by 3.1% and 7% respectively. Moreover, it is also basically optimal among the other seven indicators like Accuracy, Precision, Recall, etc., showcasing its breakthrough progress in the field of meteorological forecasting.