Anomaly Detection and Imputation Method for Residential Electricity Consumption Data Using a Temperature-Clustered Memory Network

Residential electricity consumption data, as a critical component of power market operations, directly affects the accuracy of core tasks such as load forecasting and grid dispatch. Traditional anomaly detection methods lack integration with data imputation processes and fail to leverage multi-source features to assist anomaly identification. To address these issues, this paper proposes a TCN-BiLSTM-DFM hybrid model that integrates feature extraction based on Temporal Convolutional Networks (TCN) and Bidirectional Long Short-Term Memory (BiLSTM) networks with anomaly detection based on Deep Feature Matching (DFM). The proposed method employs a cascaded TCN-BiLSTM architecture to extract features from electricity consumption and temperature time series, capturing both local abrupt changes and long-term dependencies. A deep feature memory bank based on temperature clustering is then constructed using the extracted features to achieve accurate anomaly detection. Simulation experiments are conducted using actual residential agent electricity procurement data from Hubei Province, with the London Smart Meter Dataset serving as auxiliary validation. Comparative analyses with various traditional and deep learning models demonstrate that the proposed model achieves high anomaly detection accuracy and data imputation capability, while being suitable for online deployment.