Research On The Problem Of Optimizing Combination Forecasting Model In Short-term Traffic Flow Forecasting

Presently, the forecast for short-time traffic flow of city is one of the hot and difficult points in the study of Intelligent Transportation System(ITS). In this talk, the combination forecasting model is applied to the forecasting work in ITS. By the study of linear combinations forecast theory, the robust statistics and bootstrap method are applied to the combination forecasting to find several methods to form the optimized combination weight, which has high robustness. A specific short-time traffic flow data is employed to test the new method mentioned above, and the result indicates that the accuracy of combination forecast can be improved by about 1%.The main work of this talk is presented as follows:(1) Analyse the combination forecast weights deeply, and combine the feature of short-time traffic flow data, we find the main reasons of forecast combination puzzle: the finite sample size, population distribution deviation, and the outliers.(2) The main purpose of using robust estimator is to limit the effect of"small"chang. This small change include the microvariations of all data, and the violent change of a few data, as well as the various possible between the two extremes. Though employing the robust estimation to estimateσ1,σ2, the performance of the Dickinson optimized weights can be improved better, the effect of these changes can be reduced, and the combination forecast accuracy can be improved.(3) Bootstrap method is a kind of second sampling statistics method, which is employed frequently in small sample analysis. Its aim is to simulate the unknown distribution by employing the existing small sample data on the base of mathematical statistics method. A modified bootstrap method is proposed in this talk, which can improve the effect of bootstrap method by increasing bootstrap sample length and reducing the bootstrap sample size. After the modified bootstrap method is applied in combination forecast, the influence of limited capacity of sample size will be reduced, and the combination forecast accuracy will be improved.(4) Generally, we accept that the Fisher z transformation can achieve the approximation of accumulate distribution function, and the approximation effect is good. A kind of new Sigmoid function is proposed in this talk. And we prove it can achieve better effect than Fisher z transformation in approximating the cumulative distribution function, by comparing the difference between Fisher z transformation and Sigmoid function with that and the cumulative distribution function.