A systematic methodology for development of a set of discrete-time sequence models for batch control based on historical and online operating data is presented and investigated experimentally. The modeling is based on the two independent characteristic time dimensions of batch processing, being time within the batch and the batch number. The model set is parsimoniously parametrized as a set of local, interdependent models which are estimated from data for as few as half a dozen batches. On the basis of state space models transformed from the acquired input–output model set, the asymptotic convergence of iterative learning control is combined with the closed-loop performance of model predictive control to form an optimal controller aiming to ensure reliable and reproducible operation of the batch process. This learning model predictive controller may also be used for optimizing control through optimization of the bioreactor operations model. The modeling and preliminary control performance is demonstrated on an industrial fed-batch protein cultivation production process. The presented methods lend themselves directly for application as Process Analytical Technologies.