The improvement of data selection method in data-driven model predictive control of uncertain systems

Yixuan Ni, Aoyun Ma, Dewei Li

• For unknown constrained systems with bounded disturbances, an input-mapping-based data-driven MPC with a data selection method is proposed. Based on multi-step data-driven MPC, a fully data-driven controller performance metric is designed. By incorporating a sliding window mechanism, the method selects data that effectively represents the system’s characteristics, thereby improving system control performance. The recursive feasibility and asymptotic stability are proven and a simulation result demonstrates the superior control performance of the proposed algorithm compared to existing approaches. Moreover, the algorithm uses online historical input and state data to design control input sequences and predict future states, without the need for system parameter identification.

Implementation Algorithms And Software Design Of Data-driven Model Predictive Control (master’s thesis)

Yixuan Ni, under the supervision of Dewei Li

• With the rapid development of technology, modern industrial systems have reached unprecedented levels of complexity and diversity. System scales are continually expanding, internal mechanisms are becoming increasingly intricate, and the dynamic interactions and dependencies among variables are growing more complicated. This complexity poses significant challenges to accurately constructing control system models. Moreover, model inaccuracies often result in the degradation of control performance and the reduction of stability. To address these challenges, data-driven model predictive control (MPC) has become increasingly important. By utilizing historical data to design control strategies, data-driven MPC reduces dependence on precise mathematical models, significantly improves system robustness, and effectively addresses the challenges posed by complexity and uncertainty. To combine theoretical innovations with practical application, this dissertation focuses on the implementation algorithms of data-driven MPC to address issues such as poor data quality, uncertain disturbances, and model mismatches within systems. Additionally, a distributed advanced MPC software has been developed. By integrating theory with practice, the study demonstrates the potential of data-driven MPC in real-world industrial environments.

Tools: Qt, C++, Python

• Developed a configuration interface, allowing the setup of model parameters, system constraints, and coupling relationships for large-scale systems.
• Implemented drag-and-drop functionality to configure subsystems, enhancing the software’s usability for managing interconnected processes.