Improved nonlinear predictive control performance using recurrent neural networks

Recurrent neural networks are known to have better multi-step predictive capability compared to feedforward neural networks, with the disadvantage that they are more difficult to train. This paper develops a novel recurrent neural network architecture, the structure of which allows formulation as a time varying linear model. Based on a quadruple tank challenge problem, the proposed recurrent neural network is shown to have superior performance compared to a similarly designed feedforward neural network.