A Temporal Convolutional Neural Network with Attention Mechanism for Industrial Non-Intrusive Load Monitoring

Abstract

Identifying industrial appliances can assist in demand-side management in smart grids. In this paper, a temporal convolutional neural network with attention mechanism based method is proposed for industrial non-intrusive load monitoring (NILM). First, the industrial load sequence is segmented into fixed length subsequences, and the ON-OFF states of appliances are also recorded simultaneously. Then, some segmented load subsequences are used as the input of the proposed method, and the corresponding classifiers of different appliances are trained by using the input together with the corresponding appliance states. Finally, the trained classifiers are used to classify the subsequences to be identified. This paper releases a dataset named Textile Mill Load Dataset (TMLD) that contains 30 days of load data from a textile mill. Experiments on this dataset show that the overall accuracy of the proposed method is over 88% when the load data is sampled once every one second and longer.