Dynamic Modeling of Climate Risks via Feature Entropy Weight Dimensionality Reduction and Federated Spatiotemporal Clustering

Extreme weather events triggered by climate change pose a severe threat to regional sustainable development. Traditional climate risk assessment exhibits limitations in processing high-dimensional meteorological data, capturing spatiotemporal correlations, and protecting data privacy. Based on the China Climate Physical Risk Index dataset (CCPRI, 1993–2023), this study proposes a framework entitled “Dynamic Modeling of Climate Risk Based on Feature Entropy Weight Dimensionality Reduction and Federated Spatiotemporal Clustering”: Four indicators—extreme low temperature, high temperature, rainfall, and drought—are synthesized into the Climate Physical Risk (CPR) index through feature entropy weight dimensionality reduction, which eliminates redundant features while retaining key risk information (information retention rate: 92.3%; computational complexity: O(nlogn)). By integrating federated learning with spatiotemporal clustering algorithms, cross-regional high-risk agglomeration areas are identified under the premise that original data remains within its domain.Experimental results show that the federated LSTM model achieves a coefficient of determination R2 of 0.87 for CPR prediction, representing an 18% improvement over traditional centralized models. Furthermore, it successfully identifies typical risk patterns, including persistent hotspots in South China, emerging drought risk areas in the Huang–Huai–Hai region, and stable low-risk zones on the Qinghai-Tibet Plateau. This framework effectively resolves challenges related to high-dimensional data redundancy, privacy barriers, and spatiotemporal modeling in climate risk assessment, thereby providing a scientific basis and new paradigm for climate risk management.