With the intensification of global climate change

frequent wildfires pose a serious threat to the safe operation of power systems. Addressing the challenges of complex data coupling and difficulty in characterizing causal relationships in wildfire prediction for power grid

this paper proposes a method based on causal driving and spatiotemporal graph convolutional network. Firstly

dynamic and static feature encoders are designed to extract high-dimensional spatiotemporal features from multi-source heterogeneous data

including meteorological time-series information

geospatial environment

transmission lines distribution and so on. Secondly

a causal discovery algorithm is used to identify the causal driving relationships between dynamic variables

constructing a dynamically evolving causal strength and causal time-delay graph structure. The causal strength matrix is decoupled into positive and negative adjacency matrices

and a causal-constrained graph convolutional module is constructed using time-delay information to aggregate and propagate higher-order information. Finally

the feature variables output by each module are fused and mapped to the target output space to generate the wildfire prediction results for power grid. A case study is conducted in California

USA

demonstrating that the proposed method can further explore the spatiotemporal evolution patterns of power grid wildfires

improve the accuracy and interpretability of wildfire prediction for power grid and provide support for power system disaster prevention.