Self-learning predictive control through physics-informed Machine Learning

Self-learning predictive control...

We use a combination of Machine Learning methods and building physics to generate thermal building models purely from measurement data. Compared to conventional Model Predictive Control, no manual modelling is necessary. Compared to pure Machine Learning based methods, the inclusion of building physics ensures that the model has physical behaviour and reduces the training time to 1-2 weeks.

The generated models are used in a predictive control framework, which predicts the thermal behaviour of the building for the next couple of hours, considering the weather forecast. Taking into account user preferences, occupancy schedules and operational constraints, the optimal energy input for the building is calculated every few minutes and send to the building.

...validated in buildings

We have validated the method in residential and light commercial buildings for heating and cooling operation, where it saves between 20% and 40% of energy compared to a state-of-industry controller. As the controller anticipates changes in weather, it also improves comfort, because it can act proactively.

Besides advantages in energy consumption and comfort, the method can exploit time-varying energy prices efficiently and is already ready for demand response schemes, which will become more and more relevant with more renewables entering the electricity grid.

Demonstrated at Empa in Switzerland