Competitive Multi‐Agent Graph Modelling and Fuzzy Decision Control for Thunderstorm Gale Forecasting

ABSTRACT

Thunderstorm gale forecasting can be formulated as a competitive multi‐agent spatio‐temporal decision problem, where meteorological stations, radar grid cells and convective regions interact under nonlinear, time‐varying and uncertain atmospheric conditions. To address this problem, this paper proposes a data‐driven competitive multi‐agent forecasting and fuzzy decision‐control framework for thunderstorm gale warning. First, a dynamic spatio‐temporal graph convolutional forecasting model, termed DSTGFP, is developed for multi‐station wind speed prediction. By integrating dynamic time warping, mutual information and geographical proximity, DSTGFP constructs an adaptive dynamic adjacency matrix to identify competition–cooperation interactions among station agents. Multi‐head attention and multi‐scale time–frequency feature extraction further allocate representation weights across heterogeneous spatio‐temporal patterns. Second, a multi‐source spatio‐temporal attention network, termed MSTA‐UNet, is constructed for radar echo extrapolation by fusing radar reflectivity, atmospheric dynamic variables and static terrain constraints. Finally, Kriging interpolation and fuzzy logic are introduced to construct a joint thunderstorm gale potential index, where wind‐speed and radar‐reflectivity risks are treated as soft competitive decision constraints. Experimental results show that DSTGFP achieves the best wind‐speed forecasting performance. The overall framework achieves a CSI of 0.5641 and a POD of 0.6471, demonstrating its effectiveness in modelling, analysing and controlling competitive multi‐agent interactions for severe convective weather forecasting, together with a FAR of 0.1852