Research on Lightning Prediction Based on GCN-LSTM Model Integrating Spatiotemporal Features

To overcome the limitations of spatiotemporal feature extraction that are inherent in conventional lightning warning algorithms relying solely on temporal analysis, we propose a novel prediction framework integrating a Graph Convolutional Network (GCN), Long Short-Term Memory (LSTM) architecture, and a multi-head attention mechanism. The methodology innovatively constructs station adjacency matrices based on geographical distances between meteorological monitoring stations in Qingdao, Shandong Province, China, where GCN layers capture inter-station spatial dependencies while LSTM units extract localized temporal dynamics. A dedicated multi-head attention module was developed to enable adaptive fusion of global spatiotemporal patterns, significantly enhancing lightning warning level prediction accuracy at target locations. The GCN-LSTM model achieved 93% accuracy, 59% precision, 64% recall, and a 59% F1 score. Experimental evaluation on operational meteorological data demonstrated the model’s superior performance: it achieved statistically significant accuracy improvements of 6% (p = 0.019), 3% (p = 0.026), and 2% (p = 0.03) over conventional LSTM, TGCN, and CNN-RNN baselines, respectively. Comprehensive assessments through precision–recall analysis, confusion matrix decomposition, and spatial generalizability tests confirmed the framework’s robustness. The key theoretical advancement introduced by this study lies in the synergistic coupling of graph-based spatial modeling with deep temporal sequence learning, augmented by attention-driven feature fusion—an architectural innovation addressing critical gaps in existing single-modality approaches. This methodology establishes a new paradigm for extreme weather prediction with direct applications in lightning hazard mitigation.