Online short-term multi-user load forecasting based on dynamic recognition of spatiotemporal dependencies

To fully and effectively exploit the spatiotemporal correlations among multi-user loads, this paper proposes an online short-term multi-user load forecasting method based on dynamic recognition of spatiotemporal dependencies. A hybrid graph convolutional network-bidirectional long short-term memory (GCN-BiLSTM) model is employed to capture the complex spatiotemporal relationships among multi-user loads. To enable dynamic recognition of spatiotemporal dependencies, a novel graph distance metric is developed by integrating temporal variations in load sequences with spatial correlation changes among multiple users. Temporal variations are quantified through cosine similarity, inverse cosine transformation, and complex-domain mapping to construct a weighting matrix. This matrix is then used to weight the absolute difference between graph adjacency matrices, which quantifies the changes in spatial correlations, thus forming the final graph distance metric. Based on this metric, clustering analysis is performed on the graph-structured training samples to identify typical spatiotemporal dependency patterns. During the forecasting phase, the graph distances between each input sample and these typical patterns are calculated in real time to determine the current spatiotemporal dependency pattern. When a pattern shift is detected, the model is locally fine-tuned using both the most temporally recent historical buffered samples and current-pattern historical buffered samples, thereby enabling more effective adaptive updates of the forecasting model. Case studies demonstrate that the proposed method significantly improves forecasting accuracy for multi-user loads.