Research and development of automated calibration and multi-objective optimization techniques applied to simulation of building energy models

Resumen

Building sector is responsible for 20% of total energy supplied worldwide and 40% of the final energy consumed in Europe. Therefore, it is a key sector for improving energy efficiency and reducing GHG emissions to mitigate the impact of climate change and global warming. In the context of reducing the energy demand in building sector, building energy performance simulation had arisen as a tool for predicting the energy use and thermal behaviour of buildings in design and operational phase. These simulation tools work with the so-called white-box building models. They are detailed models based on physical laws of heat and mass transfer. The advantage of using white-box models in building simulation is that they can predict building performance if changes are made to the building, like energy conservation measures. Moreover, the model inputs can be directly identified with physical parameters. The main drawback is that they are over-parametrized involving high uncertainties. Consequently, numerous studies have reported that the simulation outputs of these models have greater discrepancies respect to measured data of energy use. Model calibration involves a series of techniques that reduce the uncertainty in the model inputs by comparison of the simulation output with measured data under the same set of conditions. The aim of this thesis is to study and implement automated calibration and multi-objective optimization techniques which can increase the reliability and competitiveness of white-box models. Before calibration, sensitivity analysis techniques detect which model inputs are influential in the model error. Two global methods of sensitivity analysis were used. Morris One At a Time (MOAT) performs a qualitative rapid screening of the parameters. Bayesian Analysis of Computer Code Output (BACCO) uses a meta-model to provide a quantitative measure of each model inputs’ effect on model response based on a prior belief. Bayesian calibration was one of the selected techniques for application in this thesis. It is based on the use of the Bayes’ rule, obtaining a posterior distribution for the calibrated model inputs based on a prior belief and the use of an emulator to reduce computational time. The main advantage of Bayesian calibration is that it naturally accounts for uncertainty in the predicted values. Bayesian calibration was applied to the case study of a PASLINK test cell with high quality monitoring data to validate the results. The test cell was modelled in the software for dynamic simulation TRNSYS. A two-stage sensitivity analysis, first with MOAT, and secondly with BACCO determined the model inputs for calibration. Bayesian calibration characterized the model inputs of a calibrated model with CV(RMSE) values of the simulated temperatures lower than 3.4% for different load tests. Multi-objective optimization consists of minimizing or maximizing several objective functions. Multi-objective optimization can be used for calibrating models if the objective functions are error functions between simulated and observed data. In this thesis, multi-objective optimization with genetic algorithms has been employed for calibration of a residential micro-cogeneration model made in TRNSYS for three single-family dwellings. The calibration reduced the CV(RMSE) of the simulated temperatures and gas consumption of the model by 67.2% and 66.8% respectively. Moreover, multi-objective optimization with genetic algorithms was also applied for exergoeconomic optimization. The daily cost of the facility decreased a 24.5% and the exergy efficiency increased from 24% to 29.5%. Furthermore, the optimization achieved to reduce the overall gas consumption by 15%. A performance comparison of 60 different multi-objective optimization-based calibration approaches for white-box building models was performed. The comparison assessed which approaches provided more accurate calibration variables prediction and which error functions were robust to different calibration datasets. Among the 60 calibration approaches, eight had 70% of their Pareto predictions with percentage errors respect to the true values lower than 10%. CV(RMSE) resulted the error function used for optimization-based calibration with more accurate results to different calibration datasets.