Speaker
Description
Model predictive control (MPC) is a leading approach in modern control system design, as it effectively considers system dynamics and operational constraints and as it can anticipate behavior. However, when dealing with unknown or overly complex systems that cannot be fully described with first-principles models, can we still predict and control their behavior effectively with MPC? Recently, data-enabled predictive control (DeePC) has gained increasing attention as a promising alternative. This approach eliminates the need for prior knowledge of the control system and negates the requirement for system identification to derive an exact model. Instead, the system can be treated as a black box, allowing direct use of input and output measurement data for control purposes. But should we really disregard all existing knowledge about the system when it comes to systems that we usually know at least partially? How operational is the data-enabled method in the real world compared to a model-based method, where the model may also be identified from data using system identification?
This work provides a comprehensive discussion of these questions based on practical experience. Analyzing an utterly unknown system and conducting a competitive analysis can be quite challenging. Therefore, an omnidirectional robot designed and built at our institute is selected to generate measurement data. An experimental performance comparison for tracking a steady state is provided using two prominent predictive control approaches: data-enabled predictive control based on Willems' fundamental lemma and model predictive control utilizing a model identified with subspace identification methods. Unlike DeePC, which does not rely on a predefined model, MPC based on subspace identification methods involves constructing a state-space model of the system through input/output observations. To ensure comparability, the same measurement data is employed to build the Hankel matrix and to generate the state-space model.
Many publications focus primarily on simulated results, often neglecting critical aspects of hardware implementation, such as data collection, computational complexity, and noise management. These elements are crucial for the practical application of the methods and require further investigation. This work addresses these concerns and explores recent advancements to enhance the performance of DeePC, including regularization. The trade-offs between model complexity, computational efficiency, and control performance in hardware experiments for each approach are discussed. This comparison provides valuable insights for practitioners and researchers in selecting the most appropriate predictive control strategy for their specific applications.
