Classification Method of Load Pattern Based on Load Curve Image Information

Load pattern (LP) classification provides the foundation for demand side oriented power system operation and control research. To address the problem that the nonlinear information embedded in load curves is not mined and utilized in LP classification, this paper proposes a classification method of LP based on load curve image information. The sample set is initially classified according to the similarity, and then if sample size imbalance occurs, the subsets with fewer samples are expanded by the variational autoencoder (VAE) with long short-term memory (LSTM) layers. The optimal performance of VAE-LSTM is achieved by hyperparameter optimization, which makes the sample expansion effectively reserve the local feature of loads. The preprocessing load series is transformed into recurrence plot (RP) using phase space reconstruction to serve as image processing sample, then a load feature extraction and LP classification model is constructed combining convolutional neural network (CNN) and support vector machine (SVM). The interval in the high dimensional feature space is used as the loss function to estimate the classification effect, and the optimal classification hyperplane of SVM is built by adjusting the classification threshold and hyperplane normal vector. Also, this paper quantitatively analyzes the RP, discusses its role in discovering nonlinear characteristics of load curve and LP classification. Simulations are performed on load data of appliances in open dataset, and power consumers in an eastern Chinese city. It shows that the proposed method has exhibited satisfactory results in terms of the effectiveness of classification and the adequacy of feature extraction.