Construction Method of the Distribution Transform Load Feature Database Based on Deep Convolutional Autoencoder

Abstract

With the large-scale access of distributed resources to distribution network operation, there are more and more prosumers on the user side. It forms the basis of load prediction and demand-side management to identify different power consumption patterns and establish a typical load characteristic database according to the load data of prosumers. Therefore, a method to build a prosumer load characteristic database based on a deep convolutional autoencoder is proposed. First, the autoencoder network was used to extract the features of the load data collected to reduce the data dimension. Then, the density weight canopy algorithm was used to precluster the data after dimensionality reduction to obtain the initial clustering center and the optimal clustering number K value. The pre-clustering results were combined with the k-means algorithm for clustering, and the typical load characteristic database of prosumers was obtained. Finally, the comparison between the clustering index and the traditional k-means clustering algorithm and the improved k-means direct clustering algorithm proves that the method can effectively improve the accuracy of clustering results.