Advanced control strategies, and model predictive control (MPC) in particular, are gaining widespread interest for building energy management systems, since they can systematically save energy and/or costs with simultaneous thermal comfort improvement, as well as adapt the energy demand according to the available renewable/residual supply. The performance of any MPC strategy is dependent on the accuracy of the model describing the system and on the quality of the forecasts of disturbances, such as weather and occupant behaviour. Deviating model parameters and inaccurate disturbance predictions are shown to give rise to increased energy costs and comfort violations if not properly accounted for, and require real-time corrective actions, thereby voiding a possibly consented demand response (DR) strategy. The question on how accurate the controller model should be in order to obtain a performant optimal controller remains unanswered.
As reinforcement learning (RL) techniques are expected to bring an added value in systems subject to uncertainties, the hypothesis is that hybrid control methods, combining the strengths of MPC and RL, might lead to better control performance. A distinction can be made between (i) uncertainties that can be improved (though not fully eliminated) by adaptation through learning (e.g. model parameters, where the model remains an approximation of the real system), i.e. the parametric uncertainty (p) and (ii) uncertainties that cannot be improved (e.g. weather forecast) but that can be better dealt with in the decision-making process, i.e. the additive uncertainty (a). On-line learning/update of controller model parameters, objective function weights for conflicting objectives, controller hyperparameters, … can improve the handling of uncertainties.
This PhD will focus on two main aspects: (i) explore and integrate different modelling paradigms (black, grey and white box) in the MPC framework to evaluate the impact of model accuracy on controller performance. Analysis of the sensitivity of MPC with respect to the different modelling paradigms will be performed. The key performance indicators required for this analysis will be defined in the course of the PhD; (ii) build on existing variants of MPC such as adaptive MPC, stochastic MPCap and RL- MPC to tackle the challenge of system uncertainties. The proposed solutions will be applied in optimal control with the aim of achieving a pre-set thermal comfort level in a building. The variants of MPC developed will be compared with recent multilevel (quasi-) Monte Carlo optimization algorithms (results of another PhD running in parallel).
The PhD fellowship is granted within the collaboration between the University of Leuven and VITO.
The successful candidate will be supervised by Prof. Lieve Helsen (KU Leuven) and co-promoted by Dr. Brida Mbuwir (VITO).
For more information, please contact Dr. Brida Mbuwir: email@example.com.
How to apply?
Applications should be submitted online and should include a copy of your CV and a cover letter.
The remainder of the selection procedure is specific to the position and will be determined by the selection panel.
You can apply for this PhD vacancy no later than January 7, 2024.
More information is available on PhD | VITO.
- You hold a M.Sc. degree in electrical engineering, software engineering, computer science, mathematical engineering or data science.
- You are fluent in English, both oral and written.