Learning-based Model Predictive Control
Project Leader: Ye Pu
Primary Contact: Ye Pu (firstname.lastname@example.org)
Keywords: machine learning; optimisation; real-time control
Disciplines: Electrical & Electronic Engineering
Control, in the era of “Internet of Things” and “Cyber-physical Systems”, faces new opportunities and challenges. The huge amount of data and ever-increasing computing capability enable us to analyse systems of unprecedented scales, and to design sophisticated controllers for complex tasks. Model predictive control (MPC, an optimisation-based control technique) has been successfully applied to modern control applications due to its ability to provide high performance and to explicitly consider constraints. However, there is a gap between the MPC theory and successful implementations in practice. The MPC technique relies on accurate descriptions of system dynamics, as well as disturbances. Conventional modelling and estimation approaches fail to deal with problems when no accurate modelling assumption is available. In this project, we will make use of online learning techniques to address this problem, because learning-based approach enjoys the advantage that no particular assumptions are made on the underlying system (or data-generation) model. Furthermore, when systems exhibit fast time-varying dynamics, online methods are particularly suitable because they operate on a limited amount of data, and adapt quickly to new system dynamics. Learning-based estimation methods will find many applications in the real-world. For example, the modern power networks have started to integrate a large amount of renewable energy devices. Since renewable power generation can be easily affected by many uncertain and complex factors in the environment (eg, weather and time), the system exhibits a complex time-varying dynamics. In this project, we will develop novel approaches for Learning-based Model Predictive Control based on online learning techniques.