Dynamic Support Vector Machine Regression Based On Recurrent Strategy And Its Applications

Support vector machines with advantages of global optimum, simple structure and strong ability to promote is one of the hot fields of machine learning and pattern recognition research. The various developments of support vector machine approach in recent years, to some extent solve the effect of new information impacting the model. However, the model structure is static in nature and can not deal with the problem of real-time condition commendably. How to establish the dynamic nature of support vector machine model and apply it to system modeling and control has very important value in the area of theoretical research and application.For dynamic modeling problems, dynamic recurrent thinking is introduced to support vector machines structure, and the dynamic support vector machine model based on recurrent strategy is proposed. A specific description of the mathematical problem and the corresponding solution algorithm is given. Application of the actual industrial data can be seen that its smoothness and generalization ability is much better than the static structure of support vector machine model. In the problem of molecular weight distribution modeling and control concerning polymer, it is difficult to complete the task of modeling if only relying on dynamic support vector machine because molecular weight distribution is not only a function of time, but also a function of chain length. A compound network structure combining dynamic support vector machine with B-spline and relatively simple optimization algorithm is proposed. In which support vector machines reflects the characteristics of time and B-spline model reflects the network characteristics of space. By using the compound model, distribution function can be effectively separated in the aspect of time and space, the three-dimensional distribution function of the tracking problem is transformed into state control in the filed of time domain, and the control difficulty is greatly reduced. Conventional state feedback control strategy is promoted, and then the track control of given molecular weight distribution can be achieved. In the case of modeling imprecise and serious interference, closed-loop integral is introduced to improve the control performence.