Importance analysis and meta-model construction with correlated variables in evaluation of thermal performance of campus buildings

Statistical energy modelling & analysis of building stock is becoming mainstream in the context of city or district scale analysis of energy saving measures. A common aspect in such analyses is that there is generally a set of key explanatory variables - or the main inputs - that are statistically related to a quantity of interest (end-use energy or CO2). In the context of energy use in buildings, it is not uncommon that the explanatory variables may be correlated. However, there has been little discussion about the correlated variables in building stock research. This paper uses a set of campus buildings as a demonstrative case study to investigate the application of variable importance and meta-model construction in the case of correlated inputs when quantifying energy demand of a building stock. The variable importance analysis can identify key factors that explain energy consumption of a building stock. To this end, it is necessary to apply methods suitable for correlated inputs because the observational data (inputs) of buildings are usually correlated. For constructing statistical energy meta-models, two types of regression models are used: linear and non-parametric models. The results indicate that the linear models perform well compared to the complicated non-parametric models in this case. In addition, a simple transformation of the response, commonly used in linear regression, can improve predictive performance of both the linear and non-parametric models.