Collinearity-oriented sensitivity analysis for patterning energy factor significance in buildings

Accurate characterization of energy factor significance stands critical for building energy efficiency. Yet, previous sensitivity analyses tended to be biased in characterizing energy effects of building factors due to their incapability of handling factor interactions, categorical effects, or collinearity risks. This paper proposes a novel stochastic level-based sensitivity analysis to fully decompose energy factor significance by integrating diverse mixed effects with factor interactions under potential collinearity risks. While the former is captured via differentiative levelling, the latter is achieved through random sampling. Depending on the severity of collinearity, a selective regression using either accurate principal component regression or simple multiple linear regression is created to quantify sensitivities with a tradeoff between accuracy and computational burden. Multifactor Gaussian Mixture Models reveal factor significance hierarchy on a multi-angle sensitivity metric. The utility of presented approach is empirically validated on a case study of 289 U S. houses with high temporal consistency and clustering quality. This approach can be utilized for robust investigations of factor significance in other energy domains where categorical effects exist while addressing collinearity trap to draw reliable efficiency decisions.