A Short-Term Load Forecasting Method for Large-Scale Power Distribution Systems Based on a Novel Spatio-Temporal Neural Network

The existing body of research on load forecasting in power distribution systems has given limited consideration to the spatial relationships among nodes. To address this gap and effectively analyze the spatio-temporal heterogeneity and non-stationarity of load in distribution systems, this study introduces a novel spatio-temporal information learning architecture aimed at accurately predicting load under evolving graph structures. The proposed architecture named Meta-graph Enhanced Convolutional Recurrent Network (MegaECRN), combines a Transformer-Encoder module with a graph convolutional recurrent module. Furthermore, the architecture incorporates a dynamic topology structure learning module driven by a meta-graph repository into the graph convolutional recurrent encoder-decoder framework. To evaluate the performance of the proposed model, extensive experiments are conducted on the IEEE 8500-node test feeder, and comprehensive performance evaluations are carried out by comparing the model with baseline and state-of-the-art (SOTA) models. The results demonstrate the superior performance of our model on this dataset.