Combining model predictive control and machine learning to develop efficient smart plant technologies
Existing chemical plant optimal control can be roughly divided into model-based methods and data-based methods. The fundamental model is the result of accumulated thermodynamics and chemical reaction engineering knowledge, and provides a meaningful prior understanding about the process and a starting point for optimal controller design. However, model-plant mismatch always occurs including disturbance and uncertainty in practical processes, so there is a limitation to design a perfect optimal controller only using the prior knowledge from the fundamental model. AI (Artificial Intelligence) is a key technology for smart plant operation; it can derive unknown system dynamics or optimal strategies via machine learning based on the plant data. However, the method of designing the optimal controller of a plant using only data-based machine learning without any prior knowledge (e.g., fundamental model) requires a huge amount of data and calculations. In addition, there is a risk that the system can be damaged due to an exploring weighted policy during training. In STEP lab, we are developing smart plant technologies that combines fundamental model-based optimal control and plant data-driven machine learning including the influence of disturbance and uncertainty to efficiently complement each method's shortcomings and utilize there strengths. Specifically, we are trying to develop a control system that learns and utilizes model-plant mismatch through data-driven machine learning based on the offset-free model predictive control (MPC) framework. Unlike the existing method in which the compensation of model-plant mismatch is entirely relied on AI, the proposed scheme additionally utilizes the stabilizing property by updating the compensatory disturbance variable signal through the offset-free MPC framework. Therefore, it is expected to more efficiently improve the productivity of smart plants than the existing methods.