4. Support Vector Regression Modle for Traval Time Prediction
Factors affecting travel time, such as traffic flow, average speed, traffic flow density and lane occupancy, can be obtained by vehicle detectors . Among these data, the data reflecting the time change of traffic flow mainly includes the traffic parameters and historical average value of the past several periods of the road sections, and the spatial change of traffic flow mainly includes the current and past traffic parameters of the upper and lower reaches of the road section, and so on. There are some complex functional relations between travel time and these basic traffic parameters. When predicting travel time, we can find out the law and establish prediction model by analyzing the data of traffic basic parameters.
A. Establishment of the model.
According to the algorithm of support vector regression machine and the factors that affect the travel time, the road sections are divided into several sections according to the distribution of vehicle detectors, and one of them is selected. The traffic flow parameters of the first four working days from 6 am to 8 am are taken as input, and the travel time of the fifth day is taken as the output to establish the prediction model.
First, it is assumed that the travel time of a section of road will be predicted. According to the traffic condition of the road section and the setting of the vehicle detector, the road section is divided into i segments, i =1, 2, 3,…, m. The traffic flow and average speed of a small section of road measured by the vehicle detector during the t period are recorded as qi (t) and vi (t), respectively, and the average travel time of all vehicles passing through the whole section of the road section during the t period is set to be T (t), based on the past period I. The average travel time T (t +1), T (t +2) and so on can be predicted by the traffic parameters of a small section (here only the traffic flow) and the travel time of the whole road section to predict the average travel time T (t + 1), T (t + 2) and so on.
Secondly, the structure and parameters of kernel function and support vector machine are selected to construct the training set as followed:{(x1, y1), …, (xl, yl)}, in which y1, y2, …, yl are average travel time for the entire road over the past few periods, x1, x2, …, xl are the vector combination of the traffic state parameters of small sections and the travel time of the whole section in the past few periods.
Finally, according to the decision function constructed by the training results, the traffic state parameters and the travel time data of the small section in the current and past periods are inputted, and the average travel time in the next period (T (t+1)) is predicted by the decision function.
B. The selection of kernel functions, model structure and parameter optimization In support vector machine, there are polynomial kernel, Gauss radial basis (RBF) kernel, Sigmoid function, Fourier kernel, B-Spline kernel and so on. The corresponding kernel function can also be constructed according to specific problems. When using support vector machines to solve practical problems, the selection of appropriate kernel function is the key factor. After determining the kernel function, it is necessary to consider the optimization problem of model structure and parameters. Now it is common practice to give some parameter combinations, by grouping the known training set data, any number of sets of data, the use of a given combination of parameters to train and predict the
remaining groups of data, repeatedly calculated, according to the prediction results, take the best worthy of the optimal model structure and parameters. Parameter optimization and adjustment is one of the problems to be further studied. Heuristic intelligent optimization algorithms such as genetic algorithm and simulated annealing algorithm can be considered for optimization.
5. Conclusion
Aiming at the characteristics of bus travel and current research status this paper proposes a prediction algorithm based on support vector machine for bus station travel time by classifying public transit GPS data reasonably in different periods. And select appropriate kernel function to verify.
Finally, this paper verifies the algorithm based on actual operational data of 6 bus in Qingdao Economic Technology Development Zone, which shows that SVM model is basically consistent with actual measured data and accuracy is higher.