Identifying suitable models for the heat dynamics of buildings - Peder Bacher, Henrik Madsen 2011
@article{Bacher2011,
author = {Bacher, Peder and Madsen, Henrik},
doi = {10.1016/j.enbuild.2011.02.005},
issn = {03787788},
journal = {Energy and Buildings},
keywords = {continuous time modelling,likelihood ratio tests},
month = feb,
publisher = {Elsevier B.V.},
title = {{Identifying suitable models for the heat dynamics of buildings}},
url = {http://linkinghub.elsevier.com/retrieve/pii/S0378778811000491},
year = {2011}
}
Key Points
- Motivation is based on the expected rise of smart meters in the home,
- Models goals are: control, forecasting, description of performance,
- A hierarchy of models, which are semi-automatically selected. A human is still involved.
This paper aims to create models for the purpose of control, forecasting, and describing performance characteristics. Performance characteristics can also be though of as equivalent thermal parameters (ETPs). A hierarchy of models are presented, which are a mixture of process and data driven models. Model performance is compared using likelihood ratio tests, and then validated using statistics and physical interpretation.
Their models require frequent readings of heat consumption(?), indoor temperature(average, normalised?), ambient air temperature, and other climate variables(hopefully clarified later). The numerous models available are evaluated from the simplest onwards until no significant improvement is found. Procedure is tested upon a single storey 120 sq. m building. Data spans 6 days over winter 2009.
Models used were grey box. I'm confused as to what the difference between indoor temperature and ambient air temperature is. How do they measure this? The physical process model is then combined with a standard state based model that is a Gaussian process. Maximum likelihood estimations were carried out using a Kalman filter (not much detail was given) using a program called CTSM. Likelihoods were then compared with the simplest model to the most complex, and the larger model which gave the most significant improvement was used. However, the model selection is not a purely automated process, and a human is required to evaluate each model to ensure accuracy. Human evaluation involves looking at auto-correlation functions, and the cumulated periodogram, plots of inputs, outputs, and residuals, and an evaluation of ETPs.
The paper managed to model average house temperature, and give ETPs and how they changed given the escalating model complexity. Plots of how residuals were reduced were also shown.