Abstract: The current satellite wildfire monitoring technology for wildfires near transmission lines has high monitoring accuracy, but there are still many underreported and misreported wildfires, especially for small wildfire. To overcome the shortcomings in extracting spatio-temporal fire features, this paper proposes a wildfire identification method for transmission lines based on a deep fusion neural network with multi-scale spatio-temporal features. This algorithm first selects 12 input features to construct a multi-temporal sample library based on remote sensing data from the Himawari-8 (H-8) geostationary satellite. Then the fusion network SCNN&TLSTM is jointly constructed based on the improved feature extractor Spatial-CNN (S-CNN) and Temporal-LSTM (T-LSTM) to automatically extract the multi-scale spatio-temporal features and fuse them for the wildfire identification. A comparison with various machine learning methods is conducted to verify the effectiveness of the proposed network, where the overall accuracies of SVM, LSTM, CNN, Res-LSTM, and Res-SPP are 74.81%, 77.61%, 80.51%, 84.13% and 86.04%, respectively. However, SCNN&TLSTM can extract richer deep spatio-temporal features, achieving an overall accuracy of 90.61%. The proposed method has been applied to wildfire monitoring on transmission lines of a provincial power grid, providing reliable support to ensure the safe and stable operation of power grids.