Research On Parameter Selection Of Support Vector Machine And Its Applications In Boiler Unit

As a new machine learning method, support vector machine (SVM) is founded on the statistical learning theory (SLT). SVM has good performance both on pattern recognition and regression problems. It was and is being one focus in the field of machine learning since its emergence in 1992. SVM was applied broadly in many fields, such as pattern classification, regression, bioinformatics, finance, etc.Based on the review of SLT and SVM history, we focused on the parameter selection of SVM in this work. Although having SLT as theoretical foundation, the performance of SVM does depend on the concrete parameters during the implementations. Besides focusing on the parameter selection, we also studied the applications of SVM in the power plant boiler unit. The main work of this dissertation includes the following:1. Based on the theory and implementation of SVM, the geometry characteristics of optimal separating hyperplane was analyzed and a new method named EDSVC (equal distance support vector classifier) was presented to find the optimal separating hyperplane, which is key to SVM building. The method searches margin sample pairs locally based on equal distance criteria first and then updates the selected margin globally. The final classifier was built on the final selected margin sample pairs, which is corresponding to the support vector of standard SVM. The theory ground of EDSVC is the same as SVM in nature, so it also has the characteristics of SVM. Because the optimal separating hyperplane is built on local sample pairs, EDSVC has the characteristics of incremental learning in nature, which makes it more suitable for online learning.2. A method based on genetic algorithm (GA) with Gray coding was integrated into the training of SVM for parameter selection, the method was called GA-SVM. It was tested on two much different multi-class datasets. The results show that GA-SVM can get higher classification precision compared with the grid search method.3. Inspired by the statistical experimental design method in industrial and scientific experiments, a method based on orthogonal experimental design (OED) was used in the SVM training process for parameter selection. To improve the parameter selection efficiency, OED method was used for the numerical simulation experiments to analyse the effect of every factor here. Every time SVM training is taken as an experiment. SVM was trained according to the experiment scheme based on OED. The experiments results were analyzed to find optimal conditions during current turn experiments after the whole OED scheme was implemented. If the optimal conditions obtained during current turn can not meet the precision orother stopping criteria, the next turn experiments was implemented according to new OED scheme. This method was tested on several datasets compared with other methods. The results show that OED can improve the performance for SVM classification.4. To evaluate SVM more comprehensively, the idea of multi-object optimization (MOO) was used to analyse the performance of SVM with different parameters, including the classification precision and the cost time. A method based on multi-object particle swarm optimization (MOPSO) was presented to analyse the performance of SVM, which is called MOPSO-SVM. Take classification precision and time cost as two objects to be optimized simultaneously in this work. The results of MOO problems are not only one optimal solution, but an optimal solution set named Pareto optimal solution set (POSS). When used in some practical problems, one can select a solution from POSS combined with some other higer level information relating to the problem without retraining SVM. Several implementations on classification datasets show the POSS of parameters combinations.5. Support vector regression (SVR) was used to model the reheater steam temperature (RST) of two 300MW boiler units in power station. The fossil fuel power plants make the major part of electric power in China, how to improve the efficiency of electric generating is a key problem, especially in nowadays China. SVM was used in the power plant to improve the stability, economy and security of operating. As an example, SVR model was built based on the operating data from DCS (distributed control system) equipments to solve the problem met during the test operating. The problem is RHT abnormal, that is, the RHT difference between two sides is beyond the limit and the overall RHT is under the standard reference value, while some tube wall of reheater is above the upper bound of standard, which is dangerous and likely to result in tube failure. SVR model can find the potential relations between RHT and operating parameters, which is useful for further operating optimization. Besides the modeling of RST, SVM(SVR) can be applied in many other subsystems in power plant boiler unit.