A Short-Term Spatial-Temporal Power Load Forecasting Method Based on Temporal Graph Convolution Network
To fully explore the spatial-temporal characteristics of power load data and further improve forecasting accuracy,this paper proposes a short-term spatial-temporal load forecasting method based on a combination of graph convolutional networks and gated recurrent neural networks.Firstly,adjacency matrices and feature matrices are constructed based on the topological structure of the power network and historical load data,creating a spatial-temporal information graph for load forecasting.This transforms load forecasting into a multivariate spatial-temporal sequence prediction problem.Subsequently,graph convolutional networks are employed to explore spatial features within the spatial-temporal information graph,while gated recurrent units are utilized to learn temporal features from historical load data.The spatial-temporal load information graph is then input into the forecasting model for training,resulting in a spatial-temporal load forecasting model based on graph convolution and gated recurrent units.Finally,the proposed prediction model is verified through the real load data set of a regional power grid in Europe,and the impact of user-side distributed new energy generation on short-term load prediction was considered.Compared with a variety of typical prediction methods,the proposed prediction model improves the prediction accuracy.The method's three-step prediction accuracy reaches 98.913%,98.239%and 97.996%respectively,and it still has high accuracy when distributed new energy generation is connected where the three-step prediction accuracy reaches 98.289%,97.990%and 97.731%respectively.
short-term load forecastingtemporal and spatial characteristicsgraph convolutiongate recurrent unitdistributed new energy
万方数据
维普
