Automatic control of batch fermentation processes provides the opportunity to regulate the operation when variations in input conditions such as changes in impurity compositions in feedstock or disturbances during the run such as equipment malfunctions may cause departure from optimal reference trajectories. A simple temperature control loop or stirrer speed controller can save a 80,000 liter batch from getting ruined.
Control of batch fermentation processes can be defined as a sequence of problems. The first problem is the determination of optimal trajectories to be followed during a batch run. Given a good model, this can be cast as an open-loop optimization problem. Another approach for determining these trajectories is to extract them from historical data bases of good batches by using statistical techniques such as principal components analysis. The second problem is the low level closed-loop control of critical process variables. This may be achieved by using several single-input single-output (SISO) control loops to regulate each controlled variable by manipulating an influential manipulated variable paired with it. The third problem is higher level control that can be addressed by selecting a multi-loop or a multivariable control approach. The former necessitates the coordination of the operation of SISO loops, the latter focuses on the development of a single controller that regulates all controlled variables by all manipulated inputs. While such a controller can be built without using any low level SISO loops, practice in other areas has favored the use of SISO loops for redundancy and reliability. In that case, the multivariable controller supplies the set-points to SISO loops. The multivariable control system can be based on linear quadratic optimal control theory or model predictive control (MPC). The optimal control theory has many success stories in various fields ranging from aerospace to manufacturing and power generation. In recent years MPC has become appealing because it can handle process constraints, disturbances, and modeling errors very effectively. MPC involves the solution of a real-time constrained optimization problem at each sampling time. While this is a limiting factor, the increase of computation speed and reduction of computation cost over the years works in favor of MPC. Techniques for addressing these three problems are discussed in Chapter 7.
Was this article helpful?