As a typical locally recurrent network, the Elman neural network is widely used in nonlinear dynamic modeling and real-time control of nonlinear systems because of its dynamic performance. The structure of the Elman neural network, which is generally determined based on user experience or trial and error, plays an important role in the network’s prediction performance. Establishing a simple and optimal structure remains a problematic issue. In order to overcome this problem, the pruning Levenberg–Marquardt (PLM) training method, in which sensitivity pruning is added into its training process to obtain the optimal structure, is proposed. In the PLM method, the training process starts with an oversized structure, then the redundant hidden and context neurons are pruned and the configuration parameters are adjusted by mainly using the LM algorithm. In one pruning operation, the hidden neuron and corresponding context neuron with the least sensitivity are removed. The pruning interval, which is adaptive based on training error, is used to evaluate the success of the last pruning step and the finish of training. Further, a multi-input multi-output Elman neural network model trained by PLM method is employed for a fed-batch penicillin fermentation process. The satisfactory results indicate that the Elman model with sensitivity pruning has an appropriate structure size and performs better than the models without pruning.