Research On Building Cooling And Heating Load Prediction Based On Feature Set Construction And Machine Learning

Due to population growth and higher demand for indoor comfort,the demand for building cooling and heating load is gradually increasing.Reasonable selection of model input and effective use of model prediction algorithms are the keys to improving prediction accuracy.This study is based on the measured data of a comprehensive office and teaching building in a university in Tianjin.It compares multiple feature engineering methods to construct feature sets.Using the deep learning algorithm and boosting algorithm on the optimized feature set,an optimal ultra-short-term cooling and heating load prediction model is established.This study compares the feature sets constructed by a total of 10 types of feature engineering methods in 6 categories,and uses the deep learning algorithm DNN and the boosting algorithm Cat Boost to establish Ultra-short-term cooling and heating load prediction model.Within the scope of the author’s search,this article is the first application of Cat Boost algorithm in the field of building cooling and heating load prediction.The delay characteristics of building load and the non-linear disturbance of the building load due to the uncertain factors are considered,so the historical time value and fluctuation value of each variable are added to the feature set.In this paper,correlation analysis is used to screen the parameters related to the cooling and heating loads.Principal component analysis is used to construct the principal components of the model input parameters,eliminating the effects of multicollinearity.K-means clustering is used to highlight the structural data structure rules.Five filter algorithms remove the influence of noise signals.The discrete wavelet transform and empirical mode decomposition are used to perform hierarchical decomposition on the cooling and heating load data to filter out noise signals.Except for the PCA,the other methods improve the prediction performance to different degrees,and achieve a prediction accuracy of about 99% R-Squared.The article analyzes the influence of various variables on the prediction of cooling and heating load and gives a ranking.The results show that indoor variables represented by illuminance have a major influence on the cooling load prediction.The indoor variable represented by temperature has a major influence on the prediction result in the heating load prediction in the heating mode with a flexible operation adjustment strategy.For the cooling and heating load prediction,the cutoff values of the R-Squared of the strictest and loose minimum feature set are specified respectively.The minimum variable set corresponding to the algorithm for cooling and heating load prediction is analyzed.The results show that for the cooling and heating load prediction of Cat Boost and DNN algorithms,only 3 to 5 types of variables needed to achieve the prediction accuracy that meets the engineering needs.