Non-Intrusive Load Decomposition Based On Data Augmentation And Deep Learning

Demand-side energy management based on load monitoring technology is an important measure to promote smart grid development.The non-intrusive load decomposition(NILD)technology has received wide attention around the world due to its low cost,easy maintenance and high security.However,the load decomposition accuracy based on low-frequency sampling characteristics needs to be improved,and the increase in the number of multi-state devices and devices that states change continuously also poses a huge challenge to the traditional decomposition methods.In recent years,with the development of new technologies such as big data and artificial intelligence,how to apply the deep learning technique have attracted wide attention in various engineering areas.Considering the advantages of automatic feature extraction and strong nonlinear mapping ability by deep learning,this thesis will explore the application of deep learning to the NILD problems based on the electrical appliance low frequency power data collected by advanced measurement system such as smart meters for residential users.The main works are summarized as follows:1)To tackle the data quality problem,the data filling,data filtering and data augmentation methods are used for data preprocessing to construct an effective sample set for the further model training.The following test on the resultant sample set shows that the data augmentation in the preprocessing methods can prevent the model from overfitting,and the model which trained with the sample set constructed by data augmentation,has better performance in the decomposition problem;2)The NILD problem is modeled as a traditional value regression problem,and a 1d CNNattention neural network method is proposed to directly construct the mapping relationship between the input total power sequence samples and the output electrical sequences.This method uses three one-dimensional convolutional neural layers and three paired transposed convolutional layers and an attention calculation layer.It combines the advantages of the former two kinds of layers that are good at mining local features and the latter is good at capturing global relationship between sequence elements.Compared with the previous decomposition method,this method has faster training speed.3)In order to further improve the accuracy of decomposition based on 2),the NILD problem is modeled as a probabilistic multi-classification problem by using softmax output layer and crossentropy loss function,and a deep attention network method based on attention calculation is proposed.The research shows that this method has higher accuracy in solving the load decomposition problem,and the training speed is faster than the previous deep learning method.It performs well in the power decomposition of continuous variable state equipment and multi state equipment.