(575g) Reducing Computational Complexity of Integrated Scheduling and Control Using Machine Learning

Industrial demand response (DR) is an important tool for demand-side management, and electricity-intensive processes can support power grid stability and derive economic benefits by modulating their production in response to electricity prices. Cryogenic air separation units (ASUs) have been identified as a prime candidate for DR engagement, owing to their considerable electricity consumption and ability to safely store products in the form of cryogenic liquids. However, electricity prices in current, fast-changing markets require production scheduling decisions to be made over a time scale in which process dynamics and control become highly relevant [1].

The (first-principles) dynamic models of ASUs (and chemical process systems in general) are, however, large scale and highly nonlinear. Solving a scheduling optimization calculation while explicitly accounting for process dynamics, as represented by such models, requires more computational effort (and solution time) than acceptable in practical use. In an effort to manage the tradeoff between schedule feasibility and computational complexity, many works [2,3] assume quasi-stationary modes of operation, with additional constraints modeling the transition capabilities of the plant and its controller. Alternatively, we previously proposed system identification as a means to represent closed-loop, input-output process dynamics using low-order dynamic models, which can be embedded in scheduling calculations [4]. Further, we applied the approach to a large-scale, industrial ASU using its historical operating data [5].

In this work, we present a data-mining approach that exploits historical process data to learn a low-dimensional, latent variable representation of process dynamics for an ASU [6]. Further, we demonstrate how system identification can be performed in the latent variable space to create reduced-order models of closed-loop process behavior. We formulate an integrated scheduling and control optimization problem using the resulting models; the problem inherently features reduced computational complexity due to its low intrinsic dimensionality, and the results of the scheduling calculation demonstrate excellent economic performance.

References:

[1] Baldea, M., & Harjunkoski, I. (2014). Integrated production scheduling and process control: A systematic review. Comp. Chem. Eng., 71, 377-390.

[2] Zhang, Q., Sundaramoorthy, A., Grossmann, I. E., & Pinto, J. M. (2016). A discrete-time scheduling model for continuous power-intensive process networks with various power contracts. Comp. Chem. Eng., 84, 382-393.

[3] Mitra, S., Pinto, J. M., & Grossmann, I. E. (2014). Optimal multi-scale capacity planning for power-intensive continuous processes under time-sensitive electricity prices and demand uncertainty. Part I: Modeling. Comp. Chem. Eng., 65, 89-101.

[4] Pattison, R. C., Touretzky, C. R., Johansson, T., Harjunkoski, I., & Baldea, M. (2016). Optimal process operations in fast-changing electricity markets: framework for scheduling with low-order dynamic models and an air separation application. Ind. Eng. Chem. Res., 55(16), 4562-4584.

[5] Tsay, C., Kumar, A., Flores-Cerrillo, J., & Baldea, M. (2019). Optimal demand response scheduling of an industrial air separation unit using data-driven dynamic models. Comp. Chem. Eng.. doi: 10.1016/j.compchemeng.2019.03.022

[6] Tsay, C and Baldea, M. (2019). Learning latent variable dynamic models for integrated production scheduling and control. arXiv:1904.04796 [math.OC]