A Missing Data Imputation Method for Electricity Consumption Data Based on TCN-Attention with Mask Tokens

With the popularization of smart meters, power companies can collect massive amounts of data from users for non-invasive load detection, electricity theft detection, etc. However, due to faults in smart meters and abnormal communication, the dataset often has missing values, making it difficult for data-driven methods to be widely applied. In this paper, we propose an unsupervised learning method based on TCN-Attention to fill in customers' electricity consumption data. Firstly, we extract electricity consumption features through TCN. In order to improve the feature extraction capability, we introduce a channel attention mechanism to TCN. The attention mechanism makes TCN focus on the effective features in the data by assigning weights. Meanwhile, we use mask tokens to cover a certain proportion of data, forcing the proposed model to learn potential feature representations of the data during the training process. Finally, the features are restored to data through feedforward neural networks. An example is given to verify the effectiveness of the proposed method.