Robust Economic MPC for Solar Thermal Systems
Institution: University of British Columbia
Funding: Natural Sciences and Engineering Research Council (NSERC)
Type: PhD research
Problem
Domestic solar thermal systems face quasi-periodic disturbances: hot water demand follows daily patterns but varies unpredictably in magnitude and timing. Model predictive control (MPC) can optimize operation, but it needs an accurate disturbance forecast. Standard Gaussian process (GP) regression requires manual kernel selection and hyperparameter tuning, which breaks down when the disturbance characteristics shift over time.
What Was Novel
I developed a self-tuning kernel Gaussian process that adapts its kernel structure and hyperparameters online, without manual intervention. The GP learns the quasi-periodic structure of the disturbance (e.g., morning and evening hot water demand peaks) and feeds these predictions into an economic MPC that minimizes operating cost while maintaining comfort constraints.
The combination is novel in two ways:
- The kernel adapts online to changing disturbance patterns, unlike standard GP-MPC approaches that fix the kernel offline.
- Economic objective rather than tracking: the controller directly minimizes energy cost, not deviation from a setpoint. This matters because the optimal operating point shifts with electricity prices and solar availability.
Results
- 15% improvement in system efficiency over baseline controllers
- 10% reduction in operational costs
- Robust handling of quasi-periodic unknown disturbances without manual tuning
Publications
- Self-tuning kernel Gaussian method for predictive control systems (Journal of Process Control, 2023)
- A hybrid Gaussian process approach to robust economic model predictive control (Journal of Process Control, 2020)