Non-intrusive Load Decomposition Model Based on Deep Fusion of Multi-modal Integration

Abstract

In order to address the problems that the current non-intrusive load decomposition model based on deep learning has limited ability to model the time-dependence of long-time power consumption information, and load decomposition using the same decomposition model for devices with different load characteristics results in errors beyond the desired level. This paper proposes a non-intrusive load decomposition model, involving CNN-LSTM-TPA decomposition model and the improved SVR-VB-STCKF model. Firstly, the CNN-LSTM decomposition model was enhanced in the time-dependent modeling capacity by the temporal pattern attention (TPA) mechanism, to capture the load characteristics of the original electricity consumption information and conduct preliminary load decomposition. Secondly, the support vector regression (SVR) was used to model the nonlinear state space of the target device and the cubature Kalman filter (CKF) algorithm was modified by the improved tracking technology and variational Bayesian to create a VB-STCKF model for secondary dynamic adjustment to the preliminary decomposition results. Finally, the proposed model was verified with the public datasets (REDD and UKDALE). The verification results indicate the proposed enhanced time-dependent modeling capacity of model and dynamic adjustment to the preliminary decomposition results are obviously effective for reducing decomposition errors.

Keywords

• load decomposition
• non-intrusive load monitoring (nilm)
• energy saving and emission reduction
• temporal pattern attention (tpa)
• convolutional neural network (cnn)
• support vector regression (svr)
• kalman filter