Non-intrusive Load Monitoring Of Smart Grid Based On Deep Learning

Smart grid is a system that enables the efficient operation of contemporary power grid systems through software and hardware upgrades such as advanced sensing and measurement technologies,control methods,etc.,of which smart meters installed at households are an important part.Non-intrusive Load Monitoring,or NILM technology,refers to a technology that predicts the signals of various electrical appliances in the house from the aggregated signals of the smart meter at the entrance of the house.It is the cornerstone of software support for the smart grid system.Specifically,the task of classifying and identifying electrical signals within a short time window is called load identification,and the prediction of electrical signals within a longer time span is called load decomposition.Many research works have been carried out on the non-invasive load monitoring technology of smart grid.The identification model is mainly based on features such as V-I trajectory as input.Such features have similar shapes under similar electrical appliances,which inevitably affects the performance of classification.in the term of decomposition model,Hidden Markov and other models are often used.Although this method has a relatively fast training speed,it does not perform well in predicting the performance of periodic appliances,and the overall error is large.In order to overcome the problems of the above methods,this paper adopts deep learning technology to deal with the task of load identification and load decomposition,and proposes a new method to solve the NILM problem.The main work and contributions of the paper are as follows: 1.For the target of identifying electrical appliances in a specific short time window,it can be regarded as a classification model.Firstly,the activation current and voltage to be identified are extracted,and the non-active current is extracted according to Fryze theory to distinguish each electrical appliance.For the extracted one-dimensional time series data,two methods of Grammy angle field and recursion graph are adopted to extract its two-dimensional features.In the construction of the classifier,this paper combines the advantages of the convolutional neural network and the long short-term memory module and finally outputs the final classification through softmax.The performance of the model is compared through the metrics of classification,and it is found that it is better than the baseline.2.For the electrical power decomposition target of steady-state active power,a temporal convolutional network model based on optimization is proposed in this paper.First,the load power sampled by the main meter and the load power of the sub-meters are preprocessed to input into the model.From the perspective of model construction,this paper mainly uses the time domain convolutional network TCN,and proposes to combine the self-attention mechanism with it to improve the model performance.In addition,this paper also combines the BiGRU to further decompose the active power? and further analyzes the MAE and other indicators of the load decomposition on this basis and decomposes to obtain a specific waveform for further analysis.It is found that the model proposed in this paper performs well in the field of decomposing power.3.Further improvements to the load identification method are proposed.Since the previous identification methods only focus on the characteristics of activation current and voltage,we further analyze the characteristics of active power,reactive power,and harmonic characteristics and combine these characteristics with the characteristics of activation current for load identification.