Machine learning driven parameter identification for grey-box thermal modelling for buildings
Résumé
District-scale energy management solutions require having building models that are both fast and capable of adapting to the reality. This paper presents an approach to improve the speed and flexibility of parametrized models in building thermal modelling, with a case study on the grey-box modelling of an individual house. It consists of performing machine learning (ML) on the parameter identification step, taking into account external variables in the process. The ML model learns how the identified values of the parameters vary according to the changing conditions and uncertainties affecting the building. The trained model is then used to frequently update the RC model with appropriate parameter values during simulation on new data. During validation, this MLenhanced modelling method provides prediction accuracy levels plausibly comparable to the classical method, with large gains in execution speed, giving it the potential to be deployed to energy controllers of buildings which are less powerful than personal computers. Highlights Perform machine learning on the parameter identification step. Fast ML-driven identification on new data, suitable for MPC applications. Plausibly comparable prediction accuracy to classical grey-box modelling.
Domaines
Sciences de l'ingénieur [physics]
Origine : Fichiers produits par l'(les) auteur(s)