Based on the powerful nonlinear mapping ability of kernel learning, this paper proposed a class of kernel learning method for the short-term traffic flow forecasting. Kernel recursive least squares(KRLS) method using approximate linear dependence(ALD) technique could reduce the computational complexity and storage capacity, the KRLS method was an online kernel learning method and was suitable for training on large-scale data sets. Kernel partial least square(KPLS) method utilized the covariance between input and output variables to extract latent features. Kernel extreme learning machine(KELM) method used the kernel function to substitute for the unknown nonlinear feature mapping of the hidden layer, in addition, the output weights of the networks could also be analytically determined by using regularization least square algorithm, hence KELM method provided better generalization performance at a much faster learning speed. In order to verify the validity of the proposed kernel learning method, this paper respectively applied the employed KELM, KPLS and ALD-KRLS methods to different traffic flow forecasting instances in different area, compared to the other methods under the same conditions. Experimental results show that the proposed kernel-based learning methods have higher forecasting accuracy and improves training speed in the short-term traffic flow forecasting.

[2019-03-05] Printed Article /doc/2017.09.0916/v2