380910 MPC-Based Supervisory Control of an Off-Gas Recovery Plant with Periodic Disturbances from Parallel Batch Reactors
Chemical Vapor Deposition (CVD) is one of the typical batch processes to produce pure silicon and thin film. While the CVD reactor can produce high-purity polysilicon, it requires a long processing time and thus defies mass production. In order to increase production rate, the number of CVD reactors working simultaneously needs to be increased. In case of polysilicon production, operators replace the deposited silicon to seed filament. This makes the reactors operating in a sequential and periodic manner. The off-gas from CVD reactors needs to be further separated and recovered to recycle back to the CVD feed. Owing to the inherent discontinuous and periodic nature of operating upstream reactors, the captured off-gas cannot reach a steady-state and is continuously transported to recovery units.
This study proposes an appropriate plant-wide control structure and model-predictive control scheme that allow for the effectively separation of the off-gas with non-steady periodic feed dynamics. One of the main control objectives is to minimize the fluctuation in the recovered flowrate to reduce the equipment size and energy consumption. The target process studied in this work produces 10,000 MT polysilicon per year and the number of CVD reactors is 29. In order to construct rigorous dynamic process models in Aspen Plus and Aspen Dynamics, CVD reaction kinetic models were constructed and their parameters were identified using genetic algorithm. Proportional-integral (PI) controllers were implemented for regulation lower-level inventory levels, and a multilayer control architecture where a centralized supervisory layer based on model predictive control (MPC) minimizes the recovered flow fluctuation, were designed using MATLAB/Simulink.
Using a rigorous plant-wide dynamic model, the performance of the developed control strategy is compared with fully decentralized classical PID schemes. Discrete mode of disturbances where some of the CVD reactors are shut down is also studied, and the benefits of using multivariable MPC approach as a supervisory RTO layer are shown and analyzed.