Wind-Solar-Load Power Scenario Reduction Based on Residual Multi-Channel Convolutional Auto-Encoders

The traditional methods of generating typical scenarios are prone to information loss and feature extraction inaccuracy. Considering the characteristics of wind-solar-load data with physical individual independence, complex complementary coupling, and the quality of data dimensionality reduction, a method of generating typical wind-solar-load scenarios based on the residual multi-channel convolutional auto-encoder (RESMCAE) is proposed. With respect to data dimensionality reduction, each independent feature extraction encoder is improved by the residual module to solve the problem of gradient disappearance, reduce information loss and ensure the information integrity. The performance of the network is enhanced through integrating multi-resolution features by the decoder of multi-layer fusion structure. The KL divergence is used for optimizing the network structure, and the Adam algorithm for parameter tuning. The proposed method is verified by comparing with traditional clustering methods on Davies-Bouldin Index (DBI) and Calinski-Harabasz Index (CHI) scores. The results show that the RESMCAE is feasible.