Short-term net load prediction of distributed energy based on STGCN-Attention

Abstract

With the development of smart grids, short-term net load prediction for distributed energy is increasingly valued, net load prediction can fully explore the difference between electricity load and renewable energy output, and is the basis for energy management and optimal scheduling. Considering that the correlation between source and charge is not considered in the traditional statistical model, and the prediction accuracy is poor, this paper proposes a short-term net load prediction method for distributed energy based on the Spatio-Temporal Graph Convolution Networks (STGCNs) and Attention mechanism. First, the time series of factors such as wind power, solar power, historical characteristics of actual load, and weather environment in the integrated energy system are mapped into the data form of graph structure, and the correlation between variables is calculated using the Maximum Information Coefficient (MIC). It is used as the weighted value of the connected edges of nodes to construct the adjacency matrix and initialize the graph network. Then, STGCN is used to autonomous feature extraction of the time series. STGCN can consider the uncertainties at both ends of the source and load at the same time, and fully extract the internal relationship between the source and load variables in the distributed energy graph network. Finally, the Attention mechanism is used to strengthen the feature extraction ability of STGCN model for time series data, improving the short-term net load prediction accuracy. The experiment uses the real power data set of an integrated energy system in a region for verification. Compared with other data-driven methods, the fusion model trained and generated has higher prediction accuracy. At the same time, the trained generated model can adapt to different scale data and different prediction lengths in practice, and has strong generalization ability.