Studying the contribution of artificial intelligence to the robust optimisation of building energy management

Context and challenges

It is necessary to sharply reduce the energy consumption of the building sector. Reduction measures can relate, on the one hand, to thermal retrofitting (insulation) but also to energy management, which is the subject of this thesis. Optimisation can significantly reduce consumption peaks: in some buildings it is possible to cancel heating demand at certain hours without compromising comfort. This possibility constitutes part of the response to the prospects of the electrification of uses (heating in particular), the penetration of renewable energies, and the management of crises linked to failures in centralised production or to the international context. AI could make it possible to improve these optimal strategies by taking into account the uncertainties regarding weather forecasts and occupant behaviour.

Objectives and methods

In a predictive-control application, it is essential to assess the accuracy of the predictive model, because a prediction error can negate the benefit of the optimisation. When this model is derived from machine learning, its accuracy depends on the choice of the learning algorithm and its configuration, on the duration of the measurements and on the quantities measured. An original contribution of the thesis will therefore be to assess these various choices against a performance criterion integrating the quality of the input data, the accuracy of the predictive model in combination with the optimisation algorithm, and the uncertainties regarding weather forecasts and occupant behaviour. Solutions based on artificial intelligence will be studied, as a replacement for physical models that require a great deal of information (wall composition and other technical characteristics) that is often difficult to collect, and a certain expertise in modelling (dividing buildings into “thermal zones”, for example). Machine learning makes it possible to obtain a model from measurements already taken by building-management systems, which avoids this stage of collecting technical data and physical modelling. The computation time is reduced, which makes it possible to optimise control over a time horizon of several days operationally and to assess the consequences of uncertainties on forecasts. The value of this approach can be assessed in a case study.