Research On Non-invasive Load Monitoring Based On Improved Convolutional Neural Network

With the general improvement of residents’ living standards,household electrical equipment is also showing a trend of diversified development.These devices rely on electricity consumption to work,and the consumption of electricity is also a major concern of residents.In order to economically and efficiently monitor the amount of electricity used by various household devices,non-invasive load monitoring provides an effective solution.The research of non-invasive load decomposition based on deep learning mainly has the following problems: 1.One of the reasons for inaccurate decomposition results is that the convolution neural network is used only for load decomposition and the information of equipment operation cycle is ignored.The length of the sliding window of convolution can be increased to match the operating period to expand the network’s receptive field,but at the same time,the number of network parameters and storage will increase.Usually,gradient disappearance,explosion and network degradation will occur in deep neural networks,which will affect the accuracy of the results.2.There are many types of electrical equipment.Generally,the load decomposition of each equipment needs to train the model once.Due to the high training cost required by deep learning model,repeated training is bound to produce high cost,and this method cannot be well promoted in practical life.In view of the above questions,the research work and contributions of this paper are as follows:(1)The sensitivity field of the model is enlarged and the computational speed and model accuracy are improved.Two-way dilated convolution is introduced in this paper.On the one hand,the receptive field of the model can be increased to obtain more equipment power characteristic information.On the other hand,it can reduce the parameters of the model and its storage capacity,so as to improve the calculation speed of the model.The problem of network degradation can be solved by using batch normalization and activation function RELU.The experimental results show that the average absolute error and integrated signal error of load decomposition are smaller than those of standard deep learning,and are closer to the actual power value,and the load identification accuracy is higher after introducing the bidirectional dilated convolutional network model with residual.(2)The generalization ability of the model is enhanced.Two transfer learning schemes are introduced,namely Appliance Transfer Learning(ATL)and crossdomain transfer learning(CTL).A general model is generated for the data of different electrical equipment used in this paper,and the pre-trained model is used for other equipment or equipment in other fields,so there is no need to train each equipment data separately,which significantly saves the calculation cost.The results show that in terms of device migration and cross-domain device migration,the convolutional network model based on expansion has smaller errors than conventional convolutional load decomposition and is more suitable for largescale promotion.