Urban External Passenger Transport Demand Prediction Method Based On Machine Learning

Urban external passenger transport demand prediction is the basis of the planning and design of the passenger transportation system. The reasonable and accurate traffic demand forecast can provide data support for the location, distribution and project approachment, which can help reaching the target of not only meeting the residents’ traffic needs, but also no overinvestments. The effect of traditional forecast model basing on time inertia principle and the related factor principle is ineffective to some extent because there is a strong correlation between social and economic factors of urban external passenger transport demand and there exist phenomenon like abnormal values and missing values in the related data. Because of the gradual perfection of social statistics in the last few years, the growing accumulation of statistical data provides the basis for the research on the new method of the scholars.The research combines two excellent methods:denosing auto encoder and random forest with traffic demand forecast to ease the shortage of shallow layer machine learning method in traffic demand forecast problem. Firstly the denosing auto encoder obtaining good initialization parameters for network easing the problem of local extreme value and gradient diffusion. At the same time the active artificial random noise which force the network to reconstruct the original input in the case of the input containing noise, which makes the network’s generalization ability stronger, robustness and not easy to over fit. Besides, considering the influence factors’ correlation and time inertia window-RF method is put forward. Method’s less overfitting during two random steps: firstly random sampling from the total training sample to training decision tree, and to evaluation the generalization performance of the decision tree model by the data not extracted. Through repeating random sampling step we get many decision tree model which form a random forest. Secondly the attribute at every split node is randomly selected in a certain decision tree. The probability of the model over fitting some specific sample is greatly reduced after the double random process. In this way the generalization of forecasting model is enhanced. The urban external passenger transport demand of Peking was also set as an example for forecasting and analysis. The prediction result is of high precision. The feasibility and effectiveness of the method is well validated, it can be applied in the prediction of urban external passenger transport demand.This study focuses on the external passenger transportation demand forecast basing on machine learning method. The origin of the method, mathematical principle and how to realize method are described in detail, it would provide researchers useful reference value during undertaking work of transportation development planning for any province or city,meanwhile having positive effect in the combination of machine learning theory and transportation problem.