Full characterization of building energy factor significance by novel integrated stochastic level-based sensitivity analysis with support vector network and multivariate clustering

Effective planning and targeted implementation of building energy efficiency programs explicitly require robust knowledge on significance pattern of involving factors for precision model-based estimates of potential energy savings. Though extensively adopted for factor significance characterization, prior sensitivity analysis (SA) frameworks fail to capture typical influences attributed to commonly used categorical factors. Moreover, they often assume the assessed factors are independent which can be invalid in many cases to generate misleading outcomes. To overcome these observable defects, this paper develops a novel integrated stochastic level-based sensitivity analysis approach for exhaustively patterning effectual significance of all available building factors of varied type while incorporating diverse factor interactions on validated reliability. While the integration of categorical factors is achieved through factor-type dependent variation levelling, the random factor interactions are apprehended by stochastic sampling and nonlinear support vector network computing to accommodate collinearity trap. Multivariate clustering identifies significance pattern via a composite multi-statistic sensitivity metric derived from stochastic energy responses. Case results of 289 U.S. residential houses validate its utility. This presented approach benefits to perform reliable full sensitivity investigations and analyses under collinearity for building energy systems with verifiable robustness.