Research On Non-Intrusive Industrial Load Identification Method Based On Deep Learning

Noninvasive load monitoring is to get the bus data to grasp the internal technology,the power consumption of each device.Fine-grained information can guide the users more scientifically electricity energy consumption,reduce the waste of electricity;At the same time the power of the electric power company can master the user behavior,thereby implementing peak shift and optimize the way of power demand.At present,non-intrusive load identification have been achieved a lot of great results,but most of the results center at residential power loads,and there are few researches on load monitoring of industrial users.With the continuous progress and expand of advanced technology,deep learning algorithm has attracted more and more attention.Therefore,this article will be collected according to the actual load data set,research based on the deep learning of noninvasive industrial load identification algorithms.The main work accomplished in this paper,including:(1)The acquisition of higher dimensional load feature set for feature selection.The mutual information algorithm is adopted.First of all,according to the feature vector and the correlation between load state characterized by sorting,and then considering the correlation between characteristics of filter redundant features,reduce the redundancy of subset.Reduce the computational complexity of the model at the same time,improve the incident detection and load identification accuracy.(2)State clustering and event marking for industrial loads to be identified.In view of the defects in traditional K-means clustering algorithm,considering the distribution features of the load data,this paper studies the optimization method based on histogram peak search to save clustering time and improve the clustering accuracy.In addition,the event marking is completed based on the state sequence,and the label data set is constructed for the event detection and load identification model.(3)The convolution neural network is used to construct the event detection model.Considering the sparsity of events in industrial loads,it is impossible to use full random uniform sampling,so firstly,the threshold detection method is used to screen a large number of load data.Then,the input samples of the event detection model are constructed from the filtered data to train the model.The final test results suggest that the two-stage event detection model considering event sparsity can accurately detect most events,and its performance is better than other comparison models.(4)The convolution neural network is used to construct the load identification model.Considering the category imbalance of event samples,the SMOTE oversampling algorithm is adopted to construct the balanced dataset.Batch USES in the process of model training,at the same time the normalization and focus loss function optimization to increase the property of the model.The final test results suggest that the load identification model considering the balance of event categories has good identification effect and high identification accuracy for most events.In addition,the balanced optimization method of SMOTE oversampling and Focal loss function improves the accuracy of load identification.