Multichannel energy monitoring based on the sliding window method in an industrial environment

Abstract

The research on nonintrusive load monitoring (NILM) typically focuses on residential and commercial users. However, high-energy industrial users also require load monitoring analysis to understand their electricity consumption patterns and load operating conditions. Currently used load monitoring methods and parameters are not applicable to industrial environments. In this paper, we propose an innovative multichannel low window sequence-to-point (MLSP) method based on deep neural networks for NILM of high-energy industrial users. First, the data from each channel of the main line terminal are normalized and then aggregated into a time series, which is several times longer than the original data to expand its data volume. Then, the multichannel low window sequence-to-point method is used to select the window data of several times the length, corresponding to the active power of the target device. Finally, the discriminative features of the data are learned using a deep neural network model. The analysis of actual industrial user data shows that the lowest mean absolute error (MAE) of the original sequence-to-point method for related branches is 20.38, and the highest correlation coefficient (CC) is 90.23%. Our proposed method reduces the MAE to 11.02 and increases the CC to 97.42%. The MAE and CC indicators of all branches decreased by 25.8% and increased by 22.4%, respectively. According to Mean Absolute Percentage Error (MAPE), the performance of this model is improved by 39.6% compared with Seq2point. Compared with the original sequence-to-point method, our proposed method achieves better load monitoring results and can more accurately decompose load fluctuations, which are closer to the actual load curve.