Optimization Methods Of Non-intrusive Load Monitoring Based On Deep Neural Network

The concept of non-intrusive load monitoring(NILM)has been proposed in the face of the growing problem of energy shortage and the increasing of electricity consumption,resulting in the demand for energy efficiency and the development of smart grid technology.However,the performance of identification and coverage of the electrical type in the current NILM method has yet to be improved.The main reason is that there is no NILM algorithm to ensure that high recognition performance is compatible with all types of electrical load,while not requiring the use of high-end hardware to collect data.And it is a promising attempt to apply the rapid-developed deep neural networks technology to the NILM task.This method does not require high sampling rate,only use low-end hardware to collect data.The nature of the characteristics in electrical load can be studied by using the depth learning ability of the depth neural networks to establish the corresponding load model to carry out the load identification and decomposition.Moreover,the NILM methods based on the deep neural networks have been shown to be superior to the traditional method such as CO(combination optimization)and FHMM(factorial hidden Markov model)in some studies,which shows that using deep neural networks to optimize the performance of NILM is of high possibility.Thus,we first analyze some existing problems of the performance of NILM methods based on deep neural networks.In order to solve this problem,two basic neural network structures,dAE(denoising auto-encoder)and RNN(recurrent neural network),are selected to deal with some tasks of recognition scenarios in NILM.Based on the structure of dAE and RNN,the optimization framework based on depth neural networks is proposed for the NILM task.The optimization architecture combines the advantages of dAE network and RNN network,which uses the characteristics of filtering and denoising of dAE network as the preprocessing network structure,then utilizes the abililty of RNN network to deal with the sequential data of long-peried as the second processing step.The two targeted training of data can improve load identification accuracy and support rate requirements of load type of the NILM method.Finally,the experimental system is used to experiment with the optimization architecture to demonstrate the validity of the optimized neural network architecture for NILM.