Predictive PID Control For Nonlinear Time-delay System Based On Neural Networks

In the industrial process control, there are different time-delays for controlled objects. The existence of time-delay always results in worse control effects and lower level stabilization. Predicting the output is the key to solve time-delay problems, so predictive control becomes one of the most important control technologies for the control of time-delay system in recent years. At present, most predictive control methods are aimed at linear system, while research on nonlinear predictive control is far to enough. However, single control method can not deal with time-delay system perfectly. In this case, it is necessary to study and explore control measure integrated with intelligent theory based on the characteristic of time-delay system. Neural network not only possesses the parallel mechanism, self-learning and adaptive ability, but also can approach arbitrary complicated nonlinear system. Therefore, it is one of the hot points to combine neural network and predictive control.Firstly, the research progress of nonlinear predictive control is summed up and the idea of predictive model, roll optimization and feedback revising for predictive control are stated. Then, the research actuality of intelligent PID is expounded. At the same time, introduce a new neural network—Extreme Learning Machine, which will be applied in the modeling and identification of nonlinear system after improvement. Based on above theory, a new predictive PID control strategy is presented for the purpose of controlling nonlinear time -delay system in industrial process. This method divides the control structure into two parts: the upper part adopts Extreme Learning Machine as intelligent predictive model and the lower part takes improved Single-Neuron PID control in the controller. In this way, time-delay can be dealt with predictive control and controller parameters can be optimized by intelligent way. Applying Extreme Learning Machine as predictive model can ensure the speed and precision. Except that, improved Single-Neuron PID control inducts the performance function of general predictive control to be the object function of controller, so that the control will be farther and more precise. This method can learn very fast and obtain smaller training error, better generalization and the smallest scope of weight matrix.