Forecasting Result Synthesis Based On "Mechanism Model+Identification Model" Strategy In Traffic Flow Forecasting

The real time forecast for short-term traffic flow is the foundation of urban traffic control and guidance,which is also one of main functions of intelligent transportation systems.It's proved by studies that none of forecasting models can dominate the others in all situations. Therefore, the chief way to improve the accuracy of forecasting must be the study on strategies of forecasting.Applied"mechanism model + identification model"strategy to short-term traffic flow forecasting, studied optimal weights of linear combination forecasting, and did numerical test with observed data of one highway's three minutes short-term traffic flow.The main contents and results are:(1) The optimal weights of combined forecasting were deduced theoretically. Particularly, based on robust statistical theory, the superiority of the simple average is proved by mathematical deduction and numerical test, though it's not optimal.(2) Used technology of continuous wavelet transform and non-linear time series analysis, preliminary analyzed its predicting accuracy based on mathematical properties of the short-term traffic flow.(3) 8 forecasting results were shown based on 8 forecasting model including support vector regression model, BP neural network model, RBF neural network model, linear and quadric exponential smoothing model and time series forecasting method which have auto-regression model, two moving average models.(4) 16 groups weights coefficient were calculated based on our formulas combined forecasting weights, robust statistical techniques and exploratory data analysis techniques. Meanwhile, the correctness of our formulas combined forecasting weights was preliminarily confirmed by errors analysis and numerical test. As is shown, to get multi-result synthesis according to optimization of complicated time series specific character, multiple forecasting result merging of"mechanism model + identification model"should use neater robust statistical techniques, exploratory data analysis techniques and bootstrap techniques, including this combination forecasting method.