Application Of Neural Network Based On Stochastic Sensitivity In Smart Grid

In order to improve the operation efficiency and intelligence of power grid more effectively,machine learning related technologies are gradually applied.Among them,neural network is the most popular.Neural network can achieve the purpose of self-learning by adjusting the way and weight of a large number of internal neurons connecting with each other.In the process of neural network training,robustness should be considered,and stochastic sensitivity can quantify the sensitivity of the model to disturbance samples.Stochastic sensitivity can be used to calculate the sensitivity of the model to the disturbance samples,which can be used to calculate the noise level of the samples or measure the robussness and stability of the model.In this paper,the neural network technology based on stochastic sensitivity is applied to two subdivision fields of smart grid to further optimize two different problems.The two application areas are non-invasive load monitoring and short-term wind speed range prediction.The research focus of non-invasive load detection is to use the data of the total meter to decompose the use of electrical appliances in the whole house.In this work,it is easy to find that the switch conditions of different electrical appliances vary greatly.For example,the refrigerator is almost always on,while the coffee machine is only on a small part of the time.This involves unbalanced classification.By introducing stochastic sensitivity,this paper improves the traditional SMOTE(Synthetic Minority Oversampling Technique)oversampling learning scheme.Experimental results show that the method has good performance in many public data sets and actual non-invasive load monitoring applications.The research focus of short-term wind speed interval prediction is to construct the prediction interval for the wind speed with great volatility in the short term.Wind power generation is the most popular renewable energy,but it has great uncertainty,which makes high-quality prediction interval a challenge.The prediction interval problem has two optimization objectives.In this paper,we first extend the stochastic sensitivity to the prediction interval domain,and then construct a three objective optimization equation with two prediction interval metrics,and use NSGA-III(Non-dominated Sorting Genetic Algorithm III)for multi-objective optimization to reduce the local generalization error of the model.At the same time,compared with the previous methods,this method reduces the number of super parameters.Experimental results show that the proposed method can improve the generalization ability of the model at different confidence levels.