City road traffic system is characterized as a system of nonlinearity

uncertainty and spatial-temporal correlation

which makes traffic system parameters description and knowledge extraction difficult

and results in current short-term traffic forecast methods can not obtain satisfactory accuracy. This paper presents a hybrid multiple-kernel support vector machine model (MSVM) for conducting short-term traffic forecast. With statistical analysis of large amounts of traffic condition data samples

the proposed model not only has a capacity of recognizing and dealing with different types of input data separately

but also takes advantages of global optimization

generalization and adaptability of support vector machine. Moreover

the parameters of the hybrid model is optimized with an improved particle swarm algorithm (PSO). Aiming at the linear correlation between real time and historical traffic condition

the nonlinear correlation between real time and previous time period

and also up and downstream traffic condition

the proposed model uses a linear kernel to extract the linear pattern of traffic flow and a nonlinear kernel to map the nonlinear pattern of residuals from the linear kernel. Both the historical regularity and time-variation characteristics of city road traffic are considered in the MSVM model so as to obtain the knowledge from the influential factors of real time traffic condition in order to improve forecast accuracy. The experimental results show that the proposed model behaves satisfactory performance and robustness

and has a good potential for applications.