0308_Using support vector machines for time series prediction

Title: Using support vector machines for time series prediction 
Authors:  U. Thissena, R. van Brakela, A. P. de Weijerb, W. J. Melssena and L. M. C. Buydensa

a Laboratory of Analytical Chemistry, University of Nijmegen, Toernooiveld 1 6525 ED, Nijmegen, The Netherlands
b Teijin Twaron Research Institute, Postbus 9600, 6800 TC, Arnhem, The Netherlands

2. Summary of Paper    
Three methods for time series prediction were studied to create the possibility giving early warning as well as studying the behavior of key parameters of possible process malfunctioning. The aim of this paper is to use a support vector machine (SVM) for time series prediction. Moreover, support vector machines (SVMs), Elman recurrent neural networks, and autoregressive moving average (ARMA) models were performed and compared based on their predicting ability. The method comparison is performed on two simulated data sets and one real-world industrial data set. 
The results showed that:
1.    For the ARMA data set, the ARMA model performs best. 
2.    For the real-world data set, the SVM performed a little worse than the Elman network because the training phase of the SVM was not feasible with this relatively large data set. 
3.    ARMA model is easy and fast in use: however, its linear behavior. 
4.    RNNs are possible to use in principle model nonlinear relations but they are difficult to train or yield unstable models. Moreover, they do not lead to global or unique solution due to differences in their initial weight set. 
5.    SVMs showed well preferment for regression and time series prediction as well as they can model nonlinear relations in an efficient and stable way. In addition, the SVM is trained as a convex optimisation problem which provides a global solution.
6.    SVMs are a very good candidate for the prediction of time series because:
a.    There is a global solution 
b.    There is no overtraining due to the specific optimization procedure.
c.    In the comparison with Elman network, SVMs do not need extra validation set for performing the task.
d.    It is easy to evaluate trained SVM decision function because of the reduced number of training data that contribute to the solution.
7.    A drawback of the SVMs is the longer training time than he Elman network and the ARMA model. 

Contribution: SVMs can be a good candidate for SWRO process prediction to create the possibility giving early warning as well as studying the behavior of key parameters of the interested process.

Monruedee Moonkhum
Email: moon@gist.ac.kr

첨부 (0)