Towards automated stochastic Grey-Box Model Calibration for Heat Transfer Coefficient Inference

Abstract

An accurate on-site measurement of the intrinsic thermal performance of the building envelope is a strong leverage to promote quality management and energy performance contracting. The building sector expects on-site measurements of the Heat Transfer Coefficient to be as short as possible and highly accurate. Perturbation measurement methods analysed by RC models have been shown to be promising with accurate results within 2 days. However, stochastic RC model calibration is highly dependent on expert knowledge. In particular, model initialisation jeopardises the calibration success if not carefully chosen a priori and all the more so with second or higher order models. This paper proposes and assesses a novel automated initialisation procedure for RC models. After having brought to light that the capacitance initialisation is critical to the calibration success, the paper shows how first-order models are almost always highly identifiable. The first-order model estimated parameters then serve as initial values for all second and higher order models. This novel procedure is in 78 % of the 156 cases tested better than or similar to expert knowledge initialisation. The results are very encouraging and suggest a successful integration in a global RC model calibration and selection workflow for a fully automated process.