Batch-to-batch iterative learning control of a fed-batch fermentation process

Abstract

Recently, iterative learning control (ILC) has been used in the run-to-run control of batch processes to directly update the control trajectory. The basic idea of ILC is to update the control trajectory for a new batch run using the information from previous batch runs so that the output trajectory converges asymptotically to the desired reference trajectory. The control policy updating is calculated using linearised models around the nominal reference process input and output trajectories. The linearised models are typically identified using multiple linear regression (MLR), partial least squares (PLS) regression, or principal component regression (PCR). ILC has been shown to be a promising method to address model-plant mismatches and unknown disturbances. This work presents several improvements of batch to batch ILC strategy with applications to a simulated fed-batch fermentation process. In order to enhance the reliability of ILC, model prediction confidence is incorporated in the ILC optimization objective function. As a result of the incorporation, wide model prediction confidence bounds are penalized in order to avoid unreliable control policy updating. This method has been proven to be very effective for selected model prediction confidence bounds penalty factors. In the attempt to further improve the performance of ILC, averaged reference trajectories and sliding window techniques were introduced. To reduce the influence of measurement noise, control policy is updated on the average input and output trajectories of the past a few batches instead of just the immediate previous batch. The linearised models are re-identified using a sliding window of past batches in that the earliest batch is removed with the newest batch added to the model identification data set. The effects of various parameters were investigated for MLR, PCR and PLS method. The technique significantly improves the control performance. In model based ILC the weighting matrices, Q and R, in the objective function have a significant impact on the control performance. Therefore, in the quest to exploit the potential of objective function, adaptive weighting parameters were attempted to study the performance of batch to batch ILC with updated models. Significant improvements in the stability of the performance for all the three methods were noticed. All the three techniques suggested have established improvements either in stability, reliability and/or convergence speed. To further investigate the versatility of ILC, the above mentioned techniques were combined and the results are discussed in this thesis.