(108e) Integrating Machine-Learning Methods for Adaptive Model Predictive Control Design

Authors: 
Hassanpour, H. - Presenter, McMaster University
Corbett, B., McMaster University
Mhaskar, P., McMaster University
Automatic control systems which are widely used in the chemical process industries have faced many challenges such as nonlinearity, multi-variable interactions, and process constraints. To address these problems, advanced control strategies, in particular model predictive control (MPC), have been developed and successfully applied to the process industries (see e.g., [1]). The choice of an appropriate and reliable model is an essential task for successful development of any MPC design. Due to several challenges that exist to develop and maintain first-principles models and increasing availability of data, data-driven models and machine-learning based techniques have received significant attention in recent years. Among these approaches, artificial neural networks (ANN), especially recurrent neural networks (RNN), have been widely employed for modeling process dynamics and model-based control designs owing to their ability to capture process nonlinearity (see e.g., [2], [3], [4], [5], and [6]).

The first step for developing data-driven models is data generation. The model ability to predict the future behavior of the system can be strongly influenced by type of data which is generated under specific operation condition. Several methods consider perturbing process inputs using typically uncorrelated pseudo-random binary input signals (PRBS) (see e.g., [7]). However, applying standard uncorrelated input signals for process excitation is not sometimes possible due to the unstable nature of the process and economic and safety issues. To solve this problem, alternative methods have been proposed in the literature using closed-loop data (see e.g., [8], [9], and [10]). In the situations where data is generated under the closed-loop condition that may induce co-linearity in the input signals, the negative implications of the over-fitting problem would get exacerbated. To address this problem, the pre-processing step should be applied to training data in order to remove any co-linearity present in input signals. In this direction, principal component analysis (PCA) can be used to reject the redundancy.

Motivated by above considerations, this presentation addresses the problem of model identification in the presence of co-linearity in the input training data. A base case scenario is used for comparison where the training data is used to build an RNN model, and used directly in the MPC formulation. As expected, under MPC implementation, breaking of the correlation between the inputs leads the process to a region where the model is no longer valid, leading to poor performance. To solve this problem, two methods are proposed. In the first method, a combination of PCA and RNN is utilized to first model the process dynamics. The model is then employed in MPC implementation in a way that naturally respects the correlation observed in the training data. In the second approach, the base case RNN model is used in MPC, but a new constraint on squared prediction error (SPE) is included to ensure that the control moves remain in the same plane as the training data. Next, an approach is presented to fully exploit the performance enhancement capabilities of MPC, i.e., allow the MPC to break the input correlation seen in the training data, while not loosing model validity. To achieve this, the process is first run under MPC, albeit with slightly relaxed constraints on the SPE, thus generating richer data (while only slightly compromising on the performance). A model identified using this new data is next used in the MPC implementation to achieve superior performance.

[1] Qin, S. J., & Badgwell, T. A. (2003). A survey of industrial model predictive control technology. Control engineering practice, 11(7), 733-764.

[2] Nikravesh, M., Farell, A. E., & Stanford, T. G. (2000). Control of nonisothermal CSTR with time varying parameters via dynamic neural network control (DNNC). Chemical Engineering Journal, 76(1), 1-16.

[3] Kittisupakorn, P., Thitiyasook, P., Hussain, M. A., & Daosud, W. (2009). Neural network based model predictive control for a steel pickling process. Journal of process control, 19(4), 579-590.

[4] Sadeghassadi, M., Macnab, C. J., Gopaluni, B., & Westwick, D. (2018). Application of neural networks for optimal-setpoint design and MPC control in biological wastewater treatment. Computers & Chemical Engineering, 115, 150-160.

[5] Wu, Z., Tran, A., Rincon, D., & Christofides, P. D. (2019). Machine learning‐based predictive control of nonlinear processes. Part I: Theory. AIChE Journal, 65(11), e16729.

[6] Wu, Z., Tran, A., Rincon, D., & Christofides, P. D. (2019). Machine‐learning‐based predictive control of nonlinear processes. Part II: Computational implementation. AIChE Journal, 65(11), e16734.

[7] Shariff, H. M., Rahiman, M. H. F., & Tajjudin, M. (2013). Nonlinear system identification: Comparison between PRBS and Random Gaussian perturbation on steam distillation pilot plant. In 2013 IEEE 3rd International Conference on System Engineering and Technology, 269-274.

[8] Forssell, U., & Ljung, L. (1999). Closed-loop identification revisited. Automatica, 35(7), 1215-1241.

[9] Qin, S. J., & Ljung, L. (2003). Closed-loop subspace identification with innovation estimation. IFAC Proceedings Volumes 36(16), 861-866.

[10] Shardt, Y. A., & Huang, B. (2011). Closed-loop identification with routine operating data: Effect of time delay and sampling time. Journal of Process Control, 21(7), 997-1010.