Neural network-based control of hot-spot temperature

The control of chemical processes has demanded control engineers to come up with controllers that can cope with nonminimum-phase systems, nonlinear systems, and systems with large, variable or unknown dead times and other variable process parameters. The introduction of minimum variance controllers particularly the generalized predictive controller (GPC) has gone far in solving the above problems since the GPC is an extended horizon predictive controller which can handle the systems outlined above. In this work however, we propose the neural network GPC (NNW GPC) controller which employs the converged weights of an artificial neural network (ANN) to directly suggest the adaptive controller parameters. This approach differs from the conventional approach of using the ANN simply for projecting future process outputs. Unlike the original Clarke algorithm the NNW GPC is an inherent nonlinear estimator and therefore identifies a nonlinear system directly from plant data. The controller is easy to program and its connection to the original GPC philosophy is quite transparent. The NNW GPC controller has been employed to demonstrate the possibility of neural network control of the hot-spot in a fixed bed catalytic reactor for sulfur dioxide oxidation on crushed vanadium pentoxide catalyst