Optimal control of fed-batch fermentation processes

Optimisation of a fed-batch fermentation process typically uses the calculus of variations or Pontryagin's maximum principle to determine an optimal feed rate profile. This often results in a singular control problem and an open loop control structure. The singular feed rate is the optimal feed rate during the singular control period and is used to control the substrate concentration in the fermenter at an optimal level. This approach is supported by biological knowledge that biochemical reaction rates are controlled by the environmental conditions in the fermenter; in this case, the substrate concentration. Since an accurate neural net-based on-line estimation of the substrate concentration has recently become available and is currently employed in industry, we are therefore able to propose a method which makes use of this estimation. The proposed method divides the optimisation problem into two parts. First, an optimal substrate concentration profile which governs the biochemical reactions in the fermentation process is determined. Then a controller is designed to track the obtained optimal profile. Since the proposed method determines the optimal substrate concentration profile, the singular control problem is therefore avoided because the substrate concentration appears nonlinearly in the system equations. Also, the process is then operated in closed loop control of the substrate concentration. The proposed method is then called "closed loop optimal control". The proposed closed loop optimal control method is then compared with the open loop optimal feed rate profile method. The comparison simulations from both primary and secondary metabolite production processes show that both methods give similar performance in a case of perfect model while the closed loop optimal control provides better performance than the open loop method in a case of plant/model mismatch. The better performance of the closed loop optimal control is due to an ability to compensate for the modelling errors using feedback