Abstract: The development of deep learning provides a good foundation for processing massive load data in power systems. However, most of the existing load forecasting methods use the temporal correlation of historical load sequences to construct models without simultaneously considering the spatial coupling characteristics between neighboring nodes and the influence of external factors. Since graph convolutional neural networks have great potential in mining spatial features in power system topology, this paper proposes a multinode load forecasting method for power systems based on the fusion of attribute-enhanced multiscale spatiotemporal graph convolutional neural networks and Transformers. Firstly, the external factors are modeled as dynamic and static attributes, attribute enhancement units are designed to encode these factors, and the dynamic coupling information of loads at each node is quantified using the fast maximum mutual information system. Secondly, a multi-scale spatiotemporal graph convolutional network is used to mine the short-term spatiotemporal features among nodes, while a Transformer is used to supplement the mining of the long-term time-domain features of each node's load. Finally, the two models are fused using a gated fusion layer. The experimental results on the New York public load dataset show that the proposed method can fully exploit the spatiotemporal coupling characteristics in the multi-node load data with higher prediction accuracy and stability.