Research On Non-intrusive Load Identification Algorithm Based On Deep Neural Network

With the continuous development of new domestic smart grid technologies,residential users,as an important consumer terminal,need to apply a more flexible two-way interaction mechanism between the electricity consumption side and the electricity sales side to increase residents' awareness of smart electricity consumption.Load monitoring is an important link in analyzing the electricity consumption behavior of residents.It can carry out non-intrusive load monitoring NILM by installing an electricity monitoring device at the home without affecting the normal life of the residents.It can specifically identify each electricity load To understand the behavior and habits of residents,and provide users with a scientific and reasonable energy-saving plan to provide a technical basis,and ultimately achieve high energy utilization,reduce energy consumption and save electricity bills.Under the environment of vigorous development of deep learning,more and more scholars have been guided to widely apply the cutting-edge technology of deep neural networks to the field of load monitoring.On the basis of the above,in order to improve the overall performance of the load identification algorithm,this paper studies the optimized deep neural network model applied to NILM,and the main work done thereby can be divided into the following parts:First,the data set is preprocessed to obtain the corresponding data sample set.In order to objectively and comprehensively evaluate the deep neural network algorithm,select reasonable evaluation indicators,and use the noise reduction and reconstruction characteristics of the separate architecture of the noise reduction automatic encoder DAE and the recurrent neural network RNN.More excellent characteristics for processing long-term sequences are extended,and a load identification algorithm based on cyclic noise reduction neural network is proposed.It uses nonlinear predictive autoencoders and variable-step input features to effectively train input features,so that On the basis of predicting the state information at the next moment,the network can accept the load sequence data adaptively expanded under the unsynchronized length,so as to complete the mapping to the target load by extracting low-dimensional features.The experiment verifies that when different power characteristics are input,the optimized network is better than the single network structure in different experimental scenarios.On the basis of retaining the respective functional characteristics of the two networks,it can The load performance score has a good complementary effect,but there is still a lot of room for improvement in the identification of some appliances.According to the limitations of the cyclic noise reduction neural network identification model,a load identification algorithm for bi-gated cyclic unit Bi-GRU and attention mechanism is proposed.By optimizing the GRU structure and coding mechanism in the sequence-to-sequence model Seq2 seq,the structure is optimized.It has more excellent information extraction ability,thus enhancing the generalization ability of the model.Adding an attention mechanism to the decoding part strengthens the weighted change of the target vector of some consumers,so that the model decomposition value is closer to the true value.Through comparative experimental analysis,it is proved that the performance of the model to identify the load reaches the optimal effect.