Bayesian inference for optimising the energy refurbishment of existing buildings – BAYREB

The project aims to establish a method for diagnosing heat loss from the building envelope based on in-situ measurements. Such methods are based on the definition of an appropriate thermal model, which is then calibrated in order to reproduce the observed measurements. The parameters of the calibrated model then provide information on the desired properties: heat loss coefficients, thermal inertia, etc.
This inverse problem of thermal characterization poses a large number of scientific challenges, due to the large gap between the complexity of real transfer phenomena in a building and the simplicity of numerical models.
Using measurements already collected in different experimental cells, the project therefore carried out an in-depth study of statistical modelling for data analysis, in order to make the best use of these data by guaranteeing a reliable diagnosis. The questions are: the choice of the type of numerical model to be calibrated; the statistical learning algorithm; the possibility of disaggregating the losses; the robustness of the estimates…
The novelty of the BAYREB proposal compared to previous projects is the specific choice of the Bayesian statistical framework, and fundamental research on the capabilities and limitations of inverse methods applied to building physics. The particularity of Bayesian inference, as opposed to other inverse methods, is that it inherently performs an inverse propagation of uncertainty: missing data, sensor inaccuracy, or simplifying model assumptions, will have a direct effect on the assessment of building characteristics and their confidence intervals.

The first major outcome of the project was the proposal of a complete workflow for thermal model calibration, including validation and verification of results. This approach is necessary in order to avoid biased results in the estimation of building performance indicators.
This approach was then used to find the limitations of characterization from in situ measurements. Based on a numerical benchmark, the robustness of the method's results under variable weather conditions was verified. It was also found that the separation of heat losses by transmission and infiltration, although theoretically possible, is more complex than initially expected and requires further development.
Additional results include: the possibility to carry out the diagnosis in real time; the proposal of models allowing to take into account in the diagnosis unobserved influences on the building heat balance (latent force models). For example, it is possible to estimate the intrinsic performance of the envelope despite the presence of occupants whose activities are not explicitly known.

The main prospects of the project are twofold:
- the reduction of the measurement time required to estimate the performance of unoccupied dwellings
- statistical modelling of coupled heat and air transfer, in order to estimate the resilience of buildings during heat waves

Rouchier, S., Busser, T., Pailha, M., Piot, A., & Woloszyn, M. (2017). Hygric characterization of wood fiber insulation under uncertainty with dynamic measurements and Markov Chain Monte-Carlo algorithm. Building and Environment, 114, 129-139.
Rouchier S (2018) Solving inverse problems in building physics: An overview of guidelines for a careful and optimal use of data, Energy and Buildings, vol. 166, p. 178-195
Rouchier S, Rabouille M, Oberlé P (2018) Calibration of simplified building energy models for paramater estimation and forecasting: stochastic versus deterministic modelling, Building and Environment, vol. 134, p.181-190
Rouchier S, Jiménez MJ, Castaño S (2019) Sequential Monte Carlo for on-line parameter estimation of a lumped building energy model, Energy and Buildings (under publication)
Juricic S, Goffart J, Rouchier S, Foucquier A, Cellier N, Fraisse G. (2020) Influence of natural weather variability on the thermal characterisation of a building envelope. Soumis à Applied Energy

Rouchier, S. & Juricic, S. Structural and practical identifiability of RC models and application to the Round Robin Test Box. IEA EBC Annex 71 first expert meeting, April 26-28th 2017, Loughborough (UK)
Rouchier S, Jiménez MJ, Castaño S. Sequential Monte-Carlo for the online estimation of the Heat Loss Coefficient, 7th International Building Physics Conference, IBPC 2018, Syracuse, USA
Juricic S, Rouchier S, Foucquier A, Fraisse G, Evaluation of the physical interpretability of calibrated building model parameters, 7th International Building Physics Conference, IBPC 2018, Syracuse, USA

The BAYREB project proposes investigating the theoretical potential of Bayesian inference applied to the diagnostics of existing buildings before their refurbishment.
The largest potential for energy savings in the building sector lies in the renovation of the existing building stock. As a result, an increasing amount of research is being dedicated to encourage decision makers. Guaranteeing the performance of a building after renovation would be an efficient incitation, but is still difficult to perform in practice. A necessary condition for a cost-effective refurbishment, adapted to each specific building, is to perform detailed diagnostics of its performance prior to picking solutions: for instance, to estimate which proportion of the energy consumption is caused by air leakage, by transfer through the envelope, or by a dysfunction of the heating systems. The solution for establishing such diagnostics is to implement inverse techniques, which are able to automatically learn from in-situ measurement data in order to construct a realistic representation of the characteristics of a building.
The importance of the energy audit of existing buildings has already motivated several projects which underlined the difficulty of applying inverse methods for the identification of building properties. The overall observation of these projects is that a reliable energy audit requires a rigorous theoretical basis in order to explore its full capabilities. Moreover, previous works on the topic of building energy performance characterisation meet two main problems:
- The most common inverse methods in engineering are deterministic and do not ensure finding the global optimum in the search space. Moreover, they often only offer point estimates of the sought properties: confidence intervals on identified parameters can only be obtained by a separate forward sensitivity analysis.
- When applied to building physics, inverse methods always rely on a simplified modelling of the buildings (RC models) and may exceedingly overlook important influences on the energy performance: occupant behaviour, HVAC control strategies and coupled influences of heat, air and moisture.
While always keeping in mind the realities of the energy performance monitoring of buildings, the BAYREB project will attempt to go closer to the fundamentals of inverse techniques, in order to apply them for this purpose. The novelty of the BAYREB proposal compared to previous projects is the specific choice of the Bayesian statistical framework, and fundamental research on the capabilities and limitations of inverse methods applied to building physics.
The principle of Bayesian inference is to draw conclusions from incomplete observations of a system, and update knowledge as more data is available. Given a set of records (sensor data and inaccuracy), and some prior assumption on the model structure (expert knowledge), one can compute the probability of their causes (energy audit, envelope properties…) by means of conditional probability and the Bayes theorem. The particularity of Bayesian inference, as opposed to other inverse methods, is that it inherently performs an inverse propagation of uncertainty: missing data, sensor inaccuracy, or simplifying model assumptions, will have a direct effect on the assessment of building characteristics and their confidence intervals. A second advantage is the fact that it can be applied to any class of mathematical function, from white-box to black-box models, as it does not require the computation of sensitivity gradients.
The ambition of the BAYREB project, compared to related work on building parameter identification, is therefore twofold: to address the problem in a fully stochastic manner which will account for all model and measurement uncertainty; and to not force end users to a given building simulation software.