Reverse Uncertainty Analysis Of Building Energy Consumption Based On Bayesian Theory

As the building energy consumption accounts for more than 30%of global energy consumption and will increase further with the development of urbanization,it is necessary to implement effective and positive methods to reduce building energy consumption.The conservation renovation plan can be determined using the calibration method by obtaining the parameters of building energy models for existing buildings.However,the true uncertainty of parameters cannot be obtained by using the manual calibration methods since the parameters in building are inherently uncertain.The automatic calibration method based on the Bayesian method can fully consider the uncertainty of parameters.The likelihood function,however,is difficult to compute,which affects the application of the Bayesian method in the building energy model calibration.Therefore,the inverse uncertainty analysis method by combining the approximate Bayesian computation(ABC)and machine learning algorithm are proposed to solve the above problems.An office building is selected as the research object to evaluate the suitability of this new method.The ABC technique can obtain the posterior distributions without computing the likelihood functions.The application of machine learning algorithm can effectively alleviate the computational cost of building energy models in ABC methods by establishing the reliable machine learning models of building energy.Firstly,the metamodeling sensitivity analysis method is used to identify the key variables affecting building energy consumption.Then,the machine learning algorithm is used to obtain the monthly or hourly building energy machine learning model.Finally,the approximate Bayesian computation has been used to obtain the results of calibration,and the analysis of building retrofit can be proposed according to the results of calibration.The main conclusions are as follows:(1)The result of this case study shows that the proposed method can provide fast and reliable calibration for building energy models.In terms of the coefficient of variation of the root-mean-square error(CV(RMSE))and normalized mean bias error(NMBE),the accuracy of this building energy model calibration method is much higher than the threshold values from the ASHRAE Guideline 14-2014.For monthly and hourly data,the CV(RMSE)of the calibration results are at least 60%and 70%higher than those specified in this standard,respectively.Meanwhile,the NMBE of the calibration result are at least 40%and 60%higher than those specified in this standard,respectively,for monthly and hourly calibration.(2)According to the result of monthly building energy data,the multivariate adaptive regression model can obtain the balance between computation cost and predictive accuracy.The selected approximate Bayesian calculation methods,in descending order of accuracy of the calculation results,are neural network,local linear regression,and rejection method.(3)According to the result of hourly building energy data,the gradient boosting machine can obtain an accurate machine learning model.The accuracy of the neural network and the local linear regression is similar,while the rejection method has the worst accuracy.However,the computation time for the neural network work is more than 100 times that of the local linear regression.(4)The results of parameter calibration are used in the analysis of building retrofit.Based on the actual situation of the target building,the proposed building retrofit plan can significantly reduce the building energy consumption.In this case,after the implementation of the energy-saving renovation plan,the building can save electricity about 26.2 MWh～32.3 MWh and reduce the CO2 emission about 16.7 ton～21.3 ton.This research proposes the method of building energy model calibration to provide fast and reliable calibration result of parameters by considering parameter uncertainty.This research can guide the formulation of the building retrofit and provide an accuracy method to exist building energy analysis.In addition,this method can be widely applied in process engineering,mechanical equipment,biomedicine and another research area needed the model calibration because the proposed method in study is easy to understand and implement.