Intelligent Optimization Of The Support Vector Machine Prediction Algorithm And Applied Research

Machine Learning is one of the most advanced and intelligence-featured domains in Artificial Intelligence research. Support Vector Machine (SVM), as a new method of Machine Learning, was proposed by Vladimir Naumovich Vapnik and other scholars in 1992. Based on Statistics Learning Theory, the method seeks for optimal learning effect under limited information by actual risk minimization with structure risk minimization. Support Vector Machine has been regarded as one of new research hotspots following Artificial Neural Network.Support Vector Machine integrates technologies in maximum interval hyper-plane, Mercer kernel, convex quadratic programming, sparse solution and slack variable and makes best effect in many challenging applications.However, as a newly-rising, revolutionary technology, Support Vector Machine needs to solve some problems including model selection, parameter optimization, large sample fast training and multi-class classification. This dissertation combines the theory, method and application of Support Vector Machine together, makes a thorough study on the model selection, parameter optimization, and large sample fast training of SVM, then proposes three predict algorithms for SVM. The main work and contribution in this dissertation is as follows:1. Proposition of Chaotic-based Particle swarm optimization Support Vector Machine (CP-SVM) Algorithm. Traditional prediction algorithm on time series makes the assumption of linearity and stationary of time series to simplify characteristic of time series. However, the assumption makes obvious difference from the practical time series, causing unsatisfactory result in practical application. Although Artificial Neural Network prediction algorithms can make certain prediction on non-linear time series it can not solve problems of over-fitting and local minimum. This dissertation proposes CP-SVM algorithm by analyzing the shortage of classical time series algorithms and characteristic of chaos time series. The novelty of this algorithms lies in that it seeks for optimal model parameter by Particle Swarm Optimization (PSO) algorithm according to the time-variation characteristics of prediction model coefficient. Such algorithm, applied in complicated non-linear, non-stationary time series as electricity load, improves precision for short-term prediction.2. Proposition of Kernel principal component analysis based Quantum-behaved Particle swarm optimization algorithm of Support Vector Machines (KQP-SVM) Algorithm. The CP-SVM algorithm can acquire higher precision but with lower speed. To improve the prediction speed under permissible precision, this dissertation proposes KQP-SVM Algorithm. The novelties of KQP-SVM are as follows: it determines reasonable time series steps according to the chaotic characteristic of time series, gets characteristics of non-linear time series with kernel Principal Component Analysis, and optimizes parameter of SVM kernel function with Quantum-behaved Particle Swarm Optimization algorithm. The proposed method is then applied to short-term load prediction, and the results demonstrate that the algorithm can improve the prediction speed under permissible precision.3. Proposition of Sequential minimal optimization Wavelet Least Square Support Vector Machine (SWLS-SVM) Algorithm. The novelties of SWLS-SVM are as follows: aiming at more complicated chaos time series, it takes advantages of Wavelet analysis on high frequency detail characteristics and the high speed of Least Squares Support Vector Machine, then constructs Wavelet Least Square Support Vector Machine prediction model. With this algorithm higher prediction precision can be acquired for more complicated chaos time series. Improved sequential minimal optimization algorithm is applied to samples to improve training speed and thus yields the sparse representation of Least Square Support Vector Machine.4. Application of the SWLS-SVM algorithm. This algorithm is applied to main temperature control system of electric boiler, especially in advanced control of thermal system is accomplished typically. Experimental results show that SWLS-SVM algorithm has better performance than Artificial Neural Networks algorithm or standard Support Vector Machine algorithms in prediction.