Application Study On Predictive Control Based On Support Vector Regression

Statistics plays basic function in solving machine learning problem. But tradintional statistics mainly study progressive theory, that is to say that tradintonal statistics is based on assumption that samples are infinite. In reality, the samples that can get are finite, so some excellent learning methods cannot get statisfied result. Statistical Learning Theory is a theory that specialied in machine teaming with finite samples. Support Vector Regression is an important area of SLT.It's basical theory is VC dimension theory and structure risk minition rule, and it prossesses good characteristic and best generalization ability.SVR has good non-linear approximate ability and generalization ablity. SVR is applied to forecast data of Box-Jenkins gas furnace. The article analyses the reason why SVR is superior to the BP neural network and self-adaptive expanding neural network. At the same time the article point out the shortcoming of SVR algorithm.In order to solve the problem that support vector machines can not deal with large scale data, this paper introduces the algorithm of sequential minimal optimization (SMO) which is adapt to solve this problem ,and makes improvements based on this algorithm to increase operational speed.The identification study of support vector machine involves various courses. But most of it is off-line. The reason is that support vector machine needs long time and large memory when it processes large-scale data. In order to solve this problem, this paper proposed an algorithm named limited memory on-line identification based on support vector regression. This algorithm avoids the problem that needs large memory and meets the demand for real time.According to the good character of nonlinear approximate ability of support vector regression, apply support vector regression into internal control. This paper proposes a design method for internal control based on support vector regression (SVR-IMC): Adopt the on-line identification algorithm based on support vector regression to build system model, at the same time build inverse model off line firstly, then modify the model on line. SVR-IMC has better robust and anti-jamming ability, and has better control performance.Dynamic matrix control builds predictive model with step response which is easy to obtain in engineering practice. As model mismatch, disturbance and noise inevitably involve the modeling error, this impacts the control performance. Considers to this kind of situation, an algorithm of dynamic matrix control based on support vector regressions error compensation is proposed. Namely, build error model on-line using support vector regressions and do one-step prediction to compensate the predictive model. Thus realize the compensation for model error and strength the robustness of the algorithm.