Autoencoder-based feature extraction for power time series data considering social information
JingYI Lin, Hao Li, Chao Zhou, Wen Li, Xuesong Shao · 2023
Predicting power time series data is a crucial technology for constructing a digital and intelligent new type of power system, and feature extraction is a key prerequisite for analysis and prediction. To address the significant periodicity, susceptibility to social information factors, and high dimensionality of power data, a novel power time series feature extraction network model based on an autoencoder for social information fusion is proposed. This model employs a self-supervised method to find a bijective function that assigns each time series with a corresponding feature vector. Moreover, this model is applicable to time series data influenced by various social information. We evaluated the proposed method on electricity consumption datasets from various Indian states and compared it with several advanced methods. The comprehensive experimental results demonstrated that the method has high accuracy and stability, and can achieve state-of-the-art performance in feature extraction tasks involving multi-domain social information data.