Model-based predictive control using genetic algorithms

Abstract A predictive method of process control is developed by combining process simulators with genetic algorithm routines. The theoretical basis of the method is described and its use is illustrated for two examples involving unsteady fluid flows. In the first, a valve is to be closed as rapidly as possible without causing a storage tank to overflow. In the second, fan settings are to be chosen in a road tunnel to maintain an acceptable balance between the power consumption of the fans and pollution concentrations experienced by drivers. In both cases, the optimum control actions cannot be known in advance. Instead, the consequences of many alternative actions are explored using process simulators and the most suitable alternative is selected for implementation. The methodology is shown to be effective even when a vast range of potential actions is possible. Instead of seeking an ultimate optimum, the aim is to provide a selection of ‘good’ options from which the process operator (human or machine) can make an informed choice. A disadvantage of the method is a relatively large demand for central processing unit time, making it unsuitable for processes where very small response times are of the essence.