Calibration of simplified building energy models for parameter estimation and forecasting: Stochastic versus deterministic modelling. (15th April 2018)

• Record Type: Journal Article
• Authors: Rouchier, Simon; Rabouille, Mickaël; Oberlé, Pierre
• Is Part Of: Building and environment. Volume 134(2018)
• Page Start: 181
• Page End: 190
• Publication Date: 2018-04-15
• Subjects: Calibration -- Uncertainty -- MCMC -- Kalman filter; Buildings -- Environmental engineering -- Periodicals; Building -- Research -- Periodicals; Constructions -- Technique de l'environnement -- Périodiques; Electronic journals; 696
• DOI: 10.1016/j.buildenv.2018.02.043
• Languages: English
• ISSNs: 0360-1323

Abstract

Abstract: Due to the ill-posedness of many inverse problems, parameter estimates are often prone to a possibly large uncertainty, caused by a series of errors and approximations in the experimental and modelling work. Stochastic state-space models for time series modelling incorporate a term of process noise that represents system error; most studies on building thermal model calibration however employ deterministic models that overlook this error. This paper investigates how accounting for modelling errors affects the results of model calibration. Several simplified models are defined to simulate the indoor temperature of an experimental test cell. Some models include process noise and others do not. The parameters of each model are then learned repeatedly by using several training datasets from the test cell. The MCMC algorithm is used for training. The robustness of parameter estimates between independent trainings is evaluated. Then, the forecasting ability of the deterministic and stochastic options are compared, in terms of accuracy and robustness. Results show that stochastic modelling considerably increases the uncertainty of parameter estimates, but ensures their consistency between separate trainings, whereas deterministic models are less robust and offer a less reliable forecasting. Highlights: Stochastic and deterministic modelling are compared in a model calibration problem. The robustness of each model is assessed by using several training datasets. The MCMC algorithm is used for training. Stochastic models yield a significantly higher parameter estimation uncertainty than deterministic models. When used for prediction, calibrated stochastic models are more cautious.