An optimization approach of stochastic dynamic system via neural networks and its industry application

Neural networks can be used to solve highly nonlinear control and optimization problems relating to the determinate processes. In fact, much of the real processes are dynamic systems with noise. The authors study the modelling and optimization of a kind of stochastic dynamic system via neural networks. First, the sample patterns and the corresponding mean value patterns of the dynamic system will be used to train a multilayered backpropagation network. Then, the trained network is used to synthesize the mean value of the inputs which minimize a given stochastic objective function. The authors choose the incremental operation scheme to get optimal mean value of the input variables for overcoming the difficulty of getting entire knowledge in a time. Finally, the authors apply previous modelling and optimization methods to solve the optimal problem of a practical urea reactor. Simulation results show that this approach is effective