Time Series Online Prediction Based on Adaptive Dynamic Adjustment Kernel Recursive Least Squares Algorithm

Abstract

Kernel recursive least squares algorithm is widely employed in online prediction of time series as a kernel expend method. In the process of recursive updating, it has a lower computational complexity and a fewer storage memory. However, with the addition of new samples, the size (computational complexity) of the kernel matrix also increases. And, in the procedure of online prediction, it can’t immediately adapt to dynamically variety environment. Consequently, it is hard to fulfil the accuracy of prediction and the efficiency of prediction. At the same time, for time series online prediction, this paper presents an improved method that adaptive dynamic adjustment kernel recursive least squares (ADA-KRLS) algorithm. We adopt the dynamic update and fixed budget criteria to propose ADA-KRLS. In the proposed algorithm, the capacity of the kernel matrix is effectively limited over time, and the computational complexity is reduced. The simulation experimental of the proposed method has been given in Lorenz Multi-dimensional time series and Dalian meteorological indexes time series. The simulation experimental results show that ADA-KRLS performs better on the accuracy and the efficiency of prediction in multi-dimensional time series multi-step online prediction.