Application research of Local support vector machines in condition trend prediction of reactor coolant pump

The difficulty in parameters selection of support vector machines (SVMs), which determines the performance of SVMs, limits the application of SVMs. In this paper, a directly determination (DD) method, which combines the existing practical approach used to compute parameters ε and C with another method used to computeλ , is introduced. This method determines the values of parameters directly from analyzing training data without running SVMs training process. The results show it gets better performance than usual grid search method in terms of predicting accuracy. Moreover, it reduces the spent time to a minimum. For predicting the condition trend of reactor coolant pump (RCP), a forecasting model which combines Local SVMs, whose parameters are determined by DD method, and Time Series is used. The results of experiments show that the model is able to predict the developing trend of time series of features reflecting the pump running condition preferably.