Abstract: A double-level neural network for feedforward decoupling control is proposed for the fermentation process characterized with time-variable, nonlinear, uncertain and multivariable coupling. The first-level network aims to build a characteristic model of coupling channel on account of the information on achievable process, and realize the prediction of how the coupling affects the controlled variables; On the other hand, the second-level network aims to build an inverse characteristic model of controlling channel, thus making the compensation output and realizing the decoupling control. Testing results showed that the control performance is better in multivariable fermentation process when based on the double-level neural network feedforward decoupling, thus getting rid of the disadvantage of feedforward decoupling method that not only relies on the exact process model but highly sensitive to the slight variation of parameters.