Design And Implementation On Non-Intrusive Load Monitoring Methodology Based On Generative Adversarial Network

Energy is one of the most important materials for the development of human society.The development and utilization of energy run through the development of human society.However,with the development of society,the demand for energy is growing rapidly,and the energy crisis is becoming more and more serious.The problem of energy shortage has become one of the key factors which restrict social development.Electric energy plays an increasingly important role in industry and daily life.Therefore,electric energy saving is the top priority to promote energy-saving and consumption reduction.At the same time,the wide usage of electric energy also brings potential safety concerns.Therefore,the realization of energy-saving and safe use of electricity has become an important proposition to be solved.It is of great significance to achieve energy-saving and safety utilization of power for alleviating the energy crisis and ensuring the safety of people’s lives and property.The key lies in how to accurately perceive the power consumption behavior of electrical equipment is load monitoring.Load monitoring is usually divided into intrusive load monitoring（ILM）and non-intrusive load monitoring（NILM）.Compared with the traditional intrusive method,NILM has the characteristics of low cost and easy promotion and has become the mainstream research direction in the field of load monitoring.With the continuous improvement of the related dataset and deep learning technology,the deep learning-based NILM algorithm will be the most promising solution.However,it also has several shortcomings:not balancing the complexity of model convergence and computational cost,the preprocessing process with prior knowledge,the complex loss function,and the insufficient high-frequency information generation ability.Based on the consideration of the above problems,this thesis establishes a NILM solution based on generative adversarial networks.First,the traditional deep learning-based NILM solution can be reduced to two different computing modes:sequence to sequence（S2S）and sequence to point（S2P）,and we propose a new computing model of sequence to subsequence（S2SS）to balance the convergence complexity and computational complexity.Second,the NILM problem is creatively modeled as a sequence generation problem,and the Generative Adversarial Network（GAN）is introduced into the NILM solution.The loss function combining the discriminator output and the L1 norm is designed to avoid the tedious loss function design process.In order to further improve the low-frequency information and high-frequency information generation ability,the U-Net network structure,Instance Normalization,fully convolutional network are deeply coupled with the proposed model for collaborative optimization.In order to verify the performance of the proposed GAN-based NILM solution,experiments are designed based on UK-DALE and REFIT datasets.The results show that the accuracy of the proposed solution is better than the seq2point solution.It not only has high accuracy in the generation of low-frequency information but also can accurately restore the details in the generation of high-frequency information.In addition,the contribution of the U-Net network structure and the Instance Normalization to the model is also explored.In conclusion,the solution we proposed has high load accuracy and good application value.