Model Predictive Control of Nonlinear Distributed Parameter Systems Using Spatial Neural-Network Architectures

In this paper the distributed parameter systems comprise first-order partial differential equations coupled with ordinary differential equations. Through time and space discretization the explicit formulation of finite-difference model is constructed. Under effects of unknown disturbances and parameter uncertainties, an online learning algorithm, by virtue of the minimal output error between the system and neuro model, is developed. If only a few output measurements are available, the spatial feedforward neural-network architecture is integrated into the nondistributed predictive control framework. The stability analysis of the closed-loop control system is addressed through the discrete-time Lyapunov function approach. Two examples including a bioreactor system governed by the population balance equation and the nonisothermal tubular reactor system are used to verify the effectiveness of the proposed method.