Spatiotemporal evolution analysis of extreme weather based on adaptive graph convolutional networks

Extreme weather is intensifying and demands forecasts that capture how hazards emerge and travel. To address non-stationary dependencies and rare tails, this paper introduces an adaptive spatiotemporal graph framework. In the scheme, first a fused relational graph blends physical priors with a learned adjacency to follow changing pathways; then a multi-scale temporal encoder links weak precursors to rapid surges; finally calibrated decision layers set cost-aware thresholds and deliver coherent maps with an audit trail. Experiments show gains at six-hour lead: area under the precision-recall curve rises by 8.5% over graph WaveNet and 22% over LSTM, fractions skill at ten kilometres improves from 0.44 to 0.52, the Brier score drops from 0.134 to 0.117, and false alarms fall by about ten percent, while seasonal scores remain between 0.41 and 0.53.