Spatiotemporal feature extraction of regional building energy consumption combining residual network and convolutional attention mechanism

When dealing with complex spatiotemporal patterns affected by multi-factor coupling, there are significant limitations in the cross-layer feature fusion ability of the model, making it difficult to effectively handle the high-dimensional spatiotemporal correlations and redundant information in building energy consumption data, resulting in insufficient capture of the multi-scale dynamic changes in the spatiotemporal features of regional building energy consumption. Therefore, a method for extracting spatiotemporal features of regional building energy consumption combining Residual Network (ResNet) and convolutional attention mechanism is proposed. First, the multi-layer structure of the residual network is used to extract the temporal and spatial features of building energy consumption respectively, and skip connections and downsampling modules are used to enhance the cross-layer feature fusion ability of the model. Then, a graph attention module and a channel attention mechanism are introduced. Taking the building monitoring area as nodes and the energy consumption correlation as edges, the weight distribution of key spatiotemporal features is dynamically learned, so as to suppress redundant information and strengthen the feature expression ability. Finally, the advancement of the proposed method is verified through experiments. The experimental results show that this method can accurately capture the hourly fluctuation law and spatial heterogeneity of building energy consumption, with a prediction deviation as low as 2.41 % and an average energy saving rate of 33.20 %, significantly superior to the traditional U-Net encoder and histogram analysis method, and has good application effects.