M-DCATDGN: Enhancing load demand forecasting with an adaptive dynamic graph neural network

Electricity load demand is influenced by a multitude of factors, exhibiting considerable volatility, cyclicality, and strong spatial dependence across regions, which poses substantial challenges to the stability and operation of power systems. While widely used deep learning models, such as Recurrent Neural Networks and Convolutional Neural Networks, have advanced short-term load forecasting, they often struggle to model complex spatial correlations. Graph Neural Networks have emerged as a promising solution to this challenge; however, most existing architectures rely on static graph structures and fail to effectively capture higher-order feature interactions and dynamic spatial dependencies. In this study, we present a graph-based load forecasting model that transforms load sequences into graph signals to more effectively capture spatio-temporal dependencies. Specifically, we introduce a novel deep graph neural network called Moran’s I Based Deep Crossing Attention Dynamic Graph Neural Network (M-DCATDGN), which is designed for short-term load forecasting tasks by robustly modeling spatio-temporal correlations and higher-order feature interactions. The architecture comprises three critical components: first, a feature normalization block that normalizes input signals to address distributional shifts during inference; second, the deep crossing and squeeze-and-excitation block, which is carefully designed to capture higher-order feature interactions, assigns differential importance to time steps to ensure that the most relevant temporal features are leveraged in the forecasting process; and finally, the dynamic fusion temporal–spatial block, which utilizes multiple dynamic graph adjacency matrices guided by Local Moran’s I to capture varying levels of spatial semantic information and enhance the training of adaptive adjacency matrices. Furthermore, experiments on datasets from eight regions in New England, USA, show that M-DCATDGN achieves up to 13.4% lower root mean squared error and 17.4% lower mean absolute error than state-of-the-art baselines, demonstrating its effectiveness in robust spatio-temporal modeling for short-term load forecasting.