The neural networks used in the methods of predictive control are usually used to identify the structure of the underlying processes and then they are utilized to provide future plant outputs which are facilitated in designing a controller. It is very important to obtain accurate future predictions of the process in the predictive control systems, since the control performances are deeply concerned with the predicted values. The neural network-based models give more accurate predictions than linear models, however, the nonlinearity of neural networkd rendered is amenable to an optimization technique in obtaining a control input, which may lead to utilization of nonlinear programming (NLP) methods. However, the local minimum problem of these methods is still outstanding and needs to be resolved by some methodology such as genetic algorithm (GA).
Firstly, we consider the problem of how to design a nonlinear predictive control system which is less sensitive to the initial values of the control input based on a neural prediction model. Since most of conventional neural network-based predictive control methods are essentially based on the nonlinear progrmming methods, the local minimum problem entails depending on selection of the initial points. It deteriorates the control performance of these methods. To solve the problem, we present an predictive control method using a GA-based optimizer, noting that it is well-known that GAs can provide more chances for finging an optimal solution in the optimization problem. The input-output constraints of a plant are dealt with by introducing special type of genetic operators and a penalty function strategy. Furthermore, this control scheme is extended to multi-input, multi-output systems using partitioned prediction models. Some computer simulations are conducted to show the effectiveness of the GA-based predictive control in comparison with the adaptive GPC algorithm. And we also present a possible application of the GA-based predictive control method in the temperature control of a room.
Secondly, we investigate a special type of prediction models where the predicted outputs are related with the future control inputs linearly to avoid the appearing nonlinearity in neural network-based predictive control schemes. Based upon the linear relationship between them, we can easily derive a closed-form of control law, which is similar in form to that of GPC, and discuss about the local convergence and stability conditions of the predictive control system. Besides, the rate and amplitude constraints of the input signal are considered in a systematic way to reduce the process time required in a programming-based method. Some simulation studies show that the proposed method is superior to the adaptive GPC when used on a plant which has large unmodeled dynamics or nonlinear processes.
Lastly, the problem of adaptive predictive control systems using just one prediction model is studied. To improve the transient response of the adaptive predictive control systems, we adopt the control scheme using multiple prediction medels and a switching technique. We construct the multiple models in consideration of the configurations of a given plant, especially, we focus on the time-delay and parameter variations in the plant. Simulation results verify that the proposed control scheme gives acceptable control results in the presence of time-delay and parameter variations of the process.
