Non-intrusive Power Grid Load Monitoring Method Based On Deep Learning

Non-intrusive Load Monitoring(NILM,Non-intrusive Load Monitoring)is a method for monitoring electrical energy consumption without installing special sensors in the entire house or office building.This method is also one of the key technologies to realize smart grids and smart cities..In recent years,many excellent deep learning algorithm models have emerged in the NILM field,which has greatly improved the accuracy of non-intrusive load monitoring,but there are still many problems.This paper reproduces three classic algorithms and five deep learning models and conducts experimental analysis through NILMTK(NILM Toolkit)to compare and analyze the typical problems of current algorithm models and algorithm design ideas.The current NILM algorithm has three main problems: there is no reasonable and uniform evaluation standard for the accuracy of the algorithm;the accuracy of the identification of electrical appliances that work in multiple states for a long time,such as refrigerators,is extremely poor;for sparsely used appliances,the algorithm recall rate and accuracy The rate is too low.This thesis studies these related issues of NILM,and the main contents are as follows:Given algorithm evaluation problems in the current NILM field,almost all papers only discuss the absolute average error-index;there is no uniform and reasonable evaluation standard.This article adds four evaluation indicators F1-score,Precision,Recall,and Accuracy to NILMTK(a tool that provides a unified evaluation standard for the NILM algorithm).And the current mainstream NILM algorithm including classic edge detection algorithm,the factor is hidden Markov model and excellent algorithm of neural network model are experimentally analyzed.The experimental results clarify the three problems of the NILM algorithm: 1)The accuracy of the electrical appliance algorithm that runs for a long time and in multiple states is very poor;2)The recall rate of the NILM algorithm is not ideal,especially for sparsely used electrical appliances;3)Existing appliances The NILM neural network model cannot be trained on large data sets.At the same time,an algorithm design idea was also obtained through experiments: for long-term,multi-state working appliances,the algorithm model should pay more attention to the part of the data feature changes.For electrical appliances that work in multiple states for a long time,deep learning algorithm models are used,and the network has been optimized from three perspectives,including optimization based on activation function;optimization based on regularization function;optimization based on neural network structure.And through these methods,an improved model based on Seq2 Seq and Window GRU is designed.In the optimization experiment of Seq2 Seq,the improved Seq2 Seq LR model in this paper improves the recognition accuracy of refrigerators by 26.31% compared with the current best Seq2 Seq.Aiming at the problem of sparsely used electrical appliances algorithm recall rate and low accuracy rate,this paper builds an indoor power consumption data collection test verification platform and analyzes the active power and power factor of commonly used electrical appliances.The reason for the low recall rate of the sparse appliance algorithm is explained from the signal point of view.Finally,this paper uses a time convolutional network(TCN)to design a nilm TCN model.By rationally and flexibly designing the receptive field of the TCN model,the feature learning ability of the model is improved.Finally,the experimental comparison proves the effectiveness of the model in improving the recall rate.