Research On Short-term Prediction Of Building Cooling Load Based On Data Completeness And Feature Migration

Due to the increase of energy demand and the improvement of carbon emission policies,China is paying more and more attention to building energy conservation at this stage,and the application scenarios of building load forecasting are becoming more and more abundant.In some new or reconstructed buildings,the characteristic data related to load are relatively scarce.In this case,how to determine the characteristic variables required for load forecasting How to judge the completeness of the new data set and still get reliable prediction results through the prediction model when the data is incomplete is very important.Firstly,based on the measured data set of a public building in Tianjin,using the feature selection method,this paper studies the influence of different features on the prediction accuracy,and constructs the minimum feature set under five prediction models;On this basis,based on the measured data of two public buildings in Beijing,this paper proposes a method to determine the completeness of data features.For the case of incomplete data features,this paper also proposes a feature migration prediction model based on source domain feature mutual information extraction.This paper proposes a data feature selection method based on the combination of diffusion kernel density estimation and maximum correlation and minimum redundancy,and obtains the minimum feature set of the corresponding model in the prediction model constructed by five algorithms: neural network,support vector regression,multiple nonlinear regression,integration algorithm and deep learning algorithm.In this paper,the contribution rate of each characteristic variable to the completeness of the data set is analyzed,and the contribution degree is sorted based on this.The results show that the indoor temperature,solar radiation intensity and outdoor ambient temperature have the highest impact on the prediction accuracy in the building cooling load prediction.Then,according to the building cooling load,the strictest and loosest minimum feature set criteria are specified respectively,and the minimum variable sets in five different models are constructed.The results show that CV can be achieved by investigating 4-8 characteristic variables＿RMSE is 30% of the prediction accuracy required by the project,and provides technical support for the construction of building energy consumption monitoring platform.Based on the minimum feature set,this paper proposes an evaluation index combining feature correlation and feature distribution similarity to evaluate whether the new feature set can achieve feature completeness,which is used as the precondition for feature migration.For the case of incomplete data features,this paper proposes a feature migration method based on source domain feature selection,Firstly,the maximum correlation and minimum redundancy method is used to extract the features of the data set in the source domain,and the best feature set is constructed as the source domain data to be migrated.Then,the migration component analysis method is used to select the migrated features in the best feature set.This method solves the problem of the decline of prediction accuracy caused by incomplete data features.Compared with the traditional feature migration learning method,The prediction accuracy of this method in long-term and short-term memory artificial neural network(LSTM)algorithm is improved by more than 30%,and the CPU occupancy rate is reduced by nearly 37%.It is proved that the model has good robustness.