Active Design of Dynamic GP Models for Model Predictive Control Using Expected Improvement
Pei Sun,Junghui Chen,Lei Xie,Hongye Su
DOI: https://doi.org/10.1002/cjce.24822
2023-01-01
The Canadian Journal of Chemical Engineering
Abstract:Modelling is a basic and key requirement for model-based controlling, monitoring, or other process strategies. In non-linear model predictive control (NMPC), although data-driven models can be more easily established than first-principle ones, representative data may not be adequately included in advance to train a complete model, which is an attractive research topic. An actively improved Gaussian process (GP) model building strategy is developed, especially for incomplete models based on the idea of Bayesian optimization. The GP model can be used online as the internal model of model predictive control (MPC) directly. The model-building objective is based on the expected improvement strategy, which can exploit information gained from the currently gathered data as well as explore uncharted regions. The proposed method is a real-time design of experiments based on variance information of GP for efficient model building with insufficient initial training data for NMPC. Multi-step ahead prediction model is considered to give full play to predicting features of NPMC. Besides, a novel disturbance rejection strategy is also proposed based on GP outputs. Two simulation results, including comparisons with some traditional algorithms, are presented to demonstrate the effectiveness of the proposed method.