Optimized Local Kernel Machines for Fast Time Series Forecasting

In practices we often expect a fast learning such as real-time or online time series forecasting. However standard algorithms learning the machines from the whole data set are often time consuming. To this end, in this paper we introduce local learning strategy considering only a subset of candidates in the neighborhood of the test point and present a general form of local kernel machines for regression. To optimize these machines, based on leave-one-out errors or bounds of the kernel machines, pattern search method is adopted for model selecting. In addition, multiple-kernels are developed for performance improving. Intensive experiments on a real world electricity load forecasting have been carried out and the results demonstrate the feasibility of our methods of obtaining an improved generalization performance at a reduced computation cost.