Grouped-neural network modeling for model predictive control

A group of feed-forward neural networks (NNs), each providing the prediction of an individual process output at a future step, is used as the dynamic prediction model for the model-based predictive control (MPC) scheme in the proposed work. These NNs are parallel (independent) rather than cascaded--they are trained and implemented in parallel. Therefore, the complexity and effort in the training stage is decreased and compounded error propagation is eliminated from the prediction. A new strategy of compensating for the process-model mismatch under this grouped-NN model structure is also developed. Effectiveness of the scheme as a general nonlinear MPC is demonstrated by simulation results.