In this work, a reduced form of dynamic flux balance model based on artificial neural networks for batch and fed-batch fermentation of xylose-utilizing engineered Saccharomyces cerevisiae RWB 218 is developed. The intracellular description of carbon metabolism in the model is simulated by a multilayer-perceptron (MLP) network. First, this hybrid model is compared to the full mechanistic dynamic flux balance analysis (DFBA) in terms of accuracy and computational time regarding available experimental data on anaerobic batch cultivation. Afterwards, it is used in a model-based sequential dynamic optimization procedure in order to maximize ethanol productivity. The initial liquid volume charged in the bioreactor, the feed flow rates in aerobic and anaerobic conditions, the final batch time, and the switching time from aerobic to anaerobic conditions are considered as decision variables. Differential evolution (DE), as a robust and efficient optimization method, is employed to solve the problem for several glucose mass fractions in the feed stream. The results show that in similar conditions a small deviation in each operating parameter from its optimal value may lead to considerable decrease in ethanol productivity.