Research On Non-intrusive Load Decomposition Method Based On Deep Learning And Differential Thinking

In recent years,smart city development has become the primary concern for rational use of resources,environmental protection and improvement of social well-being.Energy management is an important part of the smart city development plan.It is necessary to manage and monitor energy consumption in buildings.The non-intrusive load monitoring technology uses only a single smart meter to monitor the operating status and energy consumption information of each power device in the building,reducing the cost of intrusive load monitoring.In this context,this article carried out in-depth research from the input signal and decomposition model of the load decomposition task,and improved the non-intrusive load decomposition system.The work is as follows:(1)In view of the insufficient information provided by the low-frequency input signal and the noise,it is difficult to classify multi-state devices.This article considers the possibility of improving the recognition performance of the system from the signal point of view.According to the idea of the change of the differential signal amplification power,the use of different reflectivity is proposed.The differential signal gives the same weight as the original input signal,changing the original one-dimensional timing input into a two-dimensional timing input,forming a new input signal mode combining the original power signal and the differential signal.Finally,the public data set is used to compare and verify the traditional deep learning model,which proves that the new signal combined with the differential signal proposed in this paper can improve the performance of the load decomposition task.(2)Aiming at the problem of slow convergence and consuming computational resources when using value regression to model the non-intrusive load decomposition,this paper uses the embedded matrix idea to convert each power value in the sequence into a high-dimensional vector after rounding and high-dimensional mapping.Therefore,the NILM task is transformed from a regression problem of obtaining the power value of each device to a probability problem of predicting the total power value at each time in each dimension of the output vector.(3)Aiming at the problem of a large amount of training parameters based on the bidirectional long short term memory model,and the problem of information compression when the input sequence is long,the paper proposes to use a bidirectional gated recurrent unit network replaces the Bi-LSTM network to reduce model parameters and subtracts model complexity;uses the attention mechanism to generate dynamic vectors that change continuously with the input to replace the fixed intermediate vector generated by the meta-model.Decompose performance for addition.This gives the final learning model of this article-an improved sequence-to-sequence combined with attention mechanism model.(4)At the end of the paper,a detailed experimental verification of the proposed new signal and S2 SA model was carried out.First,four sub-experiments were designed to select appropriate preprocessing methods and hyperparameters for the model,and then the overall performance of the model was tested in the visible and invisible scenarios.It was verified,and finally the experimental results were analyzed in detail.The experiment proved that the new signal and S2 SA model proposed in this article have excellent performance in each evaluation index.