Research On Short-term Power Load Forecasting Of Smart Grid Based On Machine Learning

With the development of smart grids,accurate power load forecasting becomes more and more important because it can help power companies better perform load scheduling and reduce excessive power production.Daily operations and planning in the smart grid require load forecasting of its customers one day in advance.However,power load forecasting is a challenging task because it relies on external factors,such as meteorology and exogenous variables.The difficulty is that there are many influencing factors and there is no way to find regular changes.This paper first explains the background and significance of short-term power load forecasting,and the research status of machine learning applied to smart grid load forecasting.Then the periodicity and holidays of the power load in an area are analyzed,and then the relationship between the load and some meteorological factors is analyzed through simulation,and several factors with high correlation that affect the load forecast are obtained.Then,in order to overcome the limitation of the current low load forecasting accuracy,the least square support vector machine(LS-SVM)model was established by preprocessing the historical load data and normalizing the related factors affecting the load forecasting.Because LS-LVM is difficult to implement training of large sample data,BP neural network model and RBF neural network model are established.Due to the inherent shortcomings of the neural network model,the final prediction results are not good,so a load prediction model based on a deep belief network(DBN)is established.The model first uses a multi-layer restricted Boltzmann machine to extract the input variables,and then predicts the load through the top-level BP neural network unit.Finally,through the experimental simulation of the load and related data in a certain area.The simulation results show that compared with the conventional neural network model,the proposed DBN model has higher prediction accuracy and the best generalization performance.