Abstract
The ability to take non-invasive Raman measurements presents a unique opportunity to use one Raman probe across multiple vessels in parallel, reducing costs but making measurements infrequent. Under these conditions, infrequent and irregular feedback signals can result in poor closed-loop control performance. This study addressed the issue of infrequent and irregular Raman measurements using a linear dynamic model developed from interpolated data to predict more frequent measurements of the controlled variable. The simulated monoclonal antibody production was sampled hourly with white noise added to the simulated glucose concentration to replicate real Raman measurements. The hourly samples were interpolated into 15 min intervals and a linear dynamic model was developed to predict the glucose concentration at 15 min intervals. These predicted values were then used in a feedback control loop by using model predictive control or a conventional proportional and integral controller to control the glucose concentration at 15 min sampling intervals. For setpoint tracking, the model predictive control reduced the integral of absolute errors to 14,600 from 15,900 (with a 1 h sampling time) or 8.2% reduction. With adaptive model predictive control, the integral of absolute errors was reduced from 14,500 (1 h sampling time) to 14,200 for setpoint tracking and from 13,500 (1 h sampling time) to 13,300 for disturbance rejection. A final comparison demonstrated that the proposed method can also cope with random variations in the sampling time.