Research On Robust Kernel Adaptive Filter Based Time Series Online Prediction

Time series are widely present in actual complex dynamic systems.So,it is of great significance to analyze and model them to excavate the dynamic behavior changes of complex systems while developing forward-looking prediction and providing auxiliary decision making.However,with the development of the era of big data,and the actual dynamic system is often in a complex noisy environment,it brings some difficulties to design online prediction methods suitable for dynamic system.Therefore,this paper takes the robust kernel adaptive filters based online time series prediction as the subject to design improved robust online prediction models,which will enhance the the ability to resist noise during the update process,reduce the time and space complexity,and further highlight the characteristics of efficiency and scalability in online prediction methods.This paper mainly studies the time series online prediction method based on the robust kernel adaptive filters,which aims to improve the robustness from the aspect of error criterion while enhancing the computational efficiency from the perspective of sparseness.The specific innovation work includes the following two contents:This paper combines the online vector quantization method with the generalized maximum correntropy criterion(GMCC),and proposes the quantized generalized maximum correntropy criterion(QGMCC),which not only increases the flexibility and generality of the similarity measurement between two variables,but also improves the computational efficiency and speed of the prediction algorithm.Then QGMCC is applied to the kernel recursive least squares(KRLS)algorithm,and the quantized kernel recursive generalized maximum correntropy(QKRGMC)algorithm is proposed,which effectively reduces the computational complexity and can better describe the time-varying characteristics of time series.Finally,in the simulations of Lorenz time series and ENSO data set,the satisfactory prediction results are achieved,which reflects the effectiveness of the model.This paper simultaneously considers the utilized of random Fourier features(RFF)method and maximum mixed correntropy criterion(MMCC)to improve the poor prediction efficiency of the kernel method.Specifically,we firstly incorporate RFF into MMCC and then this new criterion is introduced in KRLS,which is called random Fourier feature kernel recursive maximum mixture correntropy(RFF-RMMC)algorithm.The RFF is able to significantly reduce the computational complexity while the MMCC can improve the robustness of the algorithm,thereby increasing the prediction accuracy in a noisy environment.Finally,the simulation results based on the Mackey-Glass time series and the Beijing Air Quality Index show that the improved model enhances the efficiency of the prediction model while ensuring the accuracy.