Towards Robustness of Data-Driven Predictive Control for Building Energy Flexibility Applications

Identifying physics-based models of complex dynamical systems such as buildings is challenging for applications such as predictive and optimal control for demand side management in the smart grid. Data-driven predictive control using machine learning algorithms show promise as a more scalable solution when considering the greater building stock. The robustness of these algorithms for different climate data, building types, quality and quantity of data, is still not yet well understood. The objective in this study is to investigate model identification and the resultant accuracy for these various contexts using the `separation of variables' technique (DPC-En) and the consequent performance implications of the data-driven controller. The DPC-En controller is tested using a closed-loop simulation testbed of a `large office' archetype building. The results show that the technique is relatively robust to missing data and different climate types and delivers promising results using limited training data without the need for disruptive excitation measures. This work contributes to enabling a greater proportion of the diverse building stock to be utilised for demand side management by harnessing their inherent energy exibility potential.