Kernel recursive maximum total generalized correntropy for time series prediction

In order to address the problem that the predictive performance of kernel adaptive filter (KAF) is degraded in non-Gaussian and impulsive noise environments, a novel robust algorithm called kernel recursive maximum total generalized correntropy (KRMTGC) algorithm is proposed. Firstly, the system model and maximum total correntropy (MTC) are briefly introduced. Secondly, the flexible total generalized correntropy criterion is utilized to replace the MTC criterion in the kernel space, and the KRMTGC algorithm is derived in detail, which exhibits enhanced robustness against outliers and non-Gaussian noise. Moreover, to further control the infinite expansion mode of the kernel matrix of the KRMTGC algorithm, the quantized KRMTGC algorithm is proposed by using vector quantization to reduce the computational complexity. Then, the local convergence performance of the KRMTGC algorithm is analyzed. Finally, simulation results of the benchmark Rossler system and the real-world El Niño-Southern Oscillation time series prediction demonstrate that the proposed algorithm outperforms other KAF algorithms in terms of prediction speed and accuracy. © 2024 South China University of Technology. All rights reserved.