Neuro-predictive process control using online controller adaptation

A novel architecture for integrating neural networks with industrial controllers is proposed, for use in predictive control of complex process systems. In the proposed method, a conventional controller, e.g., a proportional-integral (PI) controller, is used to control the process. In addition, a recurrent neural network is used in the form of a multistep-ahead predictor (MSP) to model the process dynamics. The parameters of the PI controller are tuned by a backpropagation-through-time (BTT)-like approach using “parallel learning” to achieve acceptable regulation and stabilization of the controlled process. The advantage of such a formulation is the effective on-line adaptation of the controller parameters while the process is in operation, and the tracking of the different process operating regimes and variations. The proposed method is used in the stabilization and transient control of u-tube steam generator (UTSG) water level. Currently, available constant-gain PI controllers are unable to stabilize the UTSG at low operating powers, necessitating manual operator control. The proposed predictive controller stabilizes the process and improves its transient performance over the entire operating range. The adaptive PI controller can handle severe operational transients in the presence of significant actuator noise and process parameter drifts that could result from aging and other wear-and-tear effects.