Nonlinear Predictive Control Based On Gaussian Process

As an advanced method of computer process control, model predictive control has beenwidely used in industrial process control. As a model-based control algorithm, theperformance of model predictive control is closely related to the choice of predictive model.With the development of science and technology, however, the size and structure of theindustrial process become more massive and more complex, the traditional modelingapproach has been difficult to completely describe such systems. With the development ofmachine learning methods, the prediction model become intelligent mode, such as artificialneural networks, support vector machines, Gaussian processes. Compared with the traditionalmodeling approach, intelligent modeling approach can well describe the complicated modernlarge-scale industrial systems. For this, intelligent model-based predictive control can adapt tothe development of modern industrial technologies.As one of statistical learning methods, Gaussian process has good ability to approximatenonlinear system, while it has good generalization performance too. In addition, not only itcan provide the predictive value, but also of the variance of the predictive value, whichindicate the reliability of the predictive value. This is the characteristics which otherintelligent modeling method does not have.This dissertation first introduces the application of Gaussian process on regression andthe basic characteristics of predictive control algorithm. With the combination of the Gaussianprocess modeling and nonlinear predictive control, a new algorithm is derived. Using thisalgorithm on the common regressive nonlinear systems, we can acquire a good control effect.The experiments verify that Gaussian process has strong nonlinear curve fitting ability andgood generalization performance, so the predictive control based on Gaussian process canachieve good control effect and the output curve could track the set curve. In addition, bymeans of the Taylor expansion, we can get the linear forms of predictive function. Accordingto the linear equation of input vector, we could derive the close form solution of the predictivecontrol law, which could greatly improve the speed of the algorithm. On this basis, this paperproposed a switching control strategy based on multiple models in order to deal with thechange of model in industrial processes. From the simulation experiment, we can see that insome case this switch strategy can identify the model’s switching accurately and quickly.Finally, natural gradient method is used in online algorithm of Gaussian process and apredictive control algorithm based on online Gaussian process is derived. This algorithm is used to in simulation experiment and the result show that is has good performance.