Non-Invasive Load Decomposition Method Based on Multi-Scale TCN and Multi-Head Self-Attention Mechanism

Abstract

As a branch of the development of smart grid, non-invasive load monitoring technology is important to promoting the information granularity of users' consumption behavior, improving the efficiency of power resource utilization, and promoting the sustainable development of smart power. For improve the efficiency of feature extraction and the accuracy of load decomposition, a non-invasive load decomposition method based on multi-scale TCN and multi-head self-attention mechanism is proposed in this paper. Firstly, expansive causal convolution of multi-scale TCN is used to expand the receptor field of convolutional kernel, and residual connection and batch normalization are added to heighten the quality and efficiency of extracting deep load features. Then, a multi-scale time information encoding and embedding method is proposed to enhance the model's capability to recognize the characteristics of electricity consumption behavior. Finally, the multi-head self-attention mechanism is used to extract important load features and historical key time point information, so as to capture the power series evolution pattern and complete load decomposition. In this paper, UKDALE and REDD residential power consumption data sets are used for training and testing. Results show that the model performs well, and the accuracy of part decomposition is enhanced compared with other existing methods.