Short-Term Building Load Prediction Based On Multi-Feature Fusion And Machine Learning

The accurate prediction of building power consumption not only plays an important role in the economic and security evaluation of distribution network operation,but also has an important reference significance for the formulation of building energy saving scheme.Because the building load is affected by many factors,the prediction accuracy is difficult to improve greatly.To improve the accuracy of building load forecasting,this paper will improve the traditional building load forecasting method from two aspects.Firstly,this paper focus on the factors that affect the change of building load,analyze the correlation between meteorological factors and building load,excavate the temporal characteristics of the building load,taking the building occupancy characteristic which is closely related to the building load as one of the characteristics of the building load forecasting model,So the training set of the building load forecasting model is a high-dimensional data set with multi-feature fusion.Secondly,this paper improves the accuracy of building load prediction based on the load prediction algorithm and model,study the machine learning methods commonly used in the research field of load forecasting,and build the building load forecast model,the main machine learning model and implementation process are as follows:1)This paper established a nonparametric building load forecasting model based on maximum likelihood estimation.This paper use gaussian process regression,decision tree and random forest algorithm to build a preliminary model of building load prediction,take three model to fit the nonlinear functions of load and related influencing factors.Then,use grid search and cross validation to estimate the maximum likelihood of the important parameters of the three models.After determining the model parameters,the new test set is imported into the model for building load prediction.2)This paper established building load forecasting model based on gaussian process of maximum posterior estimation.Aim at the problem that the maximal likelihood Gaussian process regression model is prone to over-fitting,based on approximate inference algorithm of sampling,obtain the maximum a posterior estimate of the parameters of the Gaussian process regression model.And compare the prediction accuracy of different covariance functions in gaussian process regression model,select the optimal covariance function.3)This paper established building load forecasting model based on longshort term memory model.In order to solve the shortcomings of the first two models,the building load forecasting model is built by use long-short term memory model algorithm of parameterized,then adjust the model parameters according to experience,determine the final forecast model and carry out the load forecast.Through the case study can be obtained: Among the three nonparametric building load forecasting models,the accuracy of random forest and decision tree is better than gaussian regression model.The Matern32 kernel function has better learning ability in gaussian regression model,the gaussian process regression model with the maximum a posteriori estimation can effectively suppress the overfitting phenomenon,which improve the prediction accuracy.The parameterized long-short term memory model has a good fitting effect for the high-dimensional small sample data set in this paper,it is a model that have the best predict performance.When the long-short term memory model considers the building occupancy characteristics,the accuracy of prediction is improved obviously.