Comparison and application potential analysis of autoencoder-based electricity pattern mining algorithms for large-scale demand response

Studies of electricity consumption behavior patterns (ECBPs) are very important for demand-side management and emission reduction. Most of the existing ECBPs studies have limitations in data volume, algorithm performance and application potential in large-scale data scenario. Based on various autoencoders, this study proposes an ECBPs mining method with better performance and wider application potential. Two indices are constructed for model evaluations. The empirical research results based on the high-frequency power consumption data of residents show that the representation learning of ECBPs through autoencoders can not only effectively reduce the data dimension (reduced by 90%) for pattern mining but also achieve an effect no less than that of pattern mining from the original data, with the difference in pattern aggregation degree between them being only 0.003. Different autoencoders have characteristics in the representation mapping of the original data space. The results also show that the curves decoded by different autoencoders reflect different characteristics suitable for different scenarios. This study solves the dimension disaster problem of ECBPs mining in the context of large-scale data, and provides a better tool for more complex ECBPs based tasks such as multi-energy prosumers modeling and energy system optimization.