Short-term Traffic Flow Prediction Algorithm Based On K-nearest Neighbor

With the traffic congestion problem getting worse, at the same time the intelligent transportation system can effectively solve the traffic congestion and receives extensive attention. The most important part of intelligent transportation systems is the traffic flow control system. Traffic flow control system can monitor the traffic flow and coordinate traffic resources to avoid traffic congestion of large area. And the short-time traffic flow information collection, information analysis and forecasting are the key to the traffic flow control system.This paper examines several common short-term traffic flow prediction method and elaborates the practicability of solving short-term traffic flow prediction with the K neighbor nonparametric regression algorithm. This paper discusses the mathematical theory and implementation steps of K neighbor nonparametric regression prediction algorithm, and then applies it to forecast taxis and passenger traffic flow in the Beijing Capital International Airport.As a nonparametric method, K neighbor nonparametric regression prediction algorithm has some problems limiting its practical application. The main defects of K neighbor nonparametric regression prediction algorithm are: because of many influence factors of traffic flow and deficiency of the pretreatment of the raw data, state vector component is too simple or too complex, which causes the unreasonable structure of historical database and eventually leads to the matching process taking too much. Due to the raw data without classification, all traffic patterns use a constant K, which cause the miscarriage of the current traffic patterns and increase the prediction error. In order to overcome these defects, this paper puts forward some measures to improve algorithm.The main measures to improve are:(1) the introduction of principal component analysis and cluster analysis to process the raw data, principal component analysis algorithm can reduce the dimension of the state vector, while also eliminating the correlation between the state vector components. Cluster analysis to classify the raw data and put the raw data which is belong to the same class together to change the uneven state of the raw data density.(2) creating a history data base with a fast search capabilities. First, the raw data、clustering center data and neighborhood data store separately. At the same time the center data of the cluster data are mapped into a one-dimensional data, and store separately in order to improve search efficiency.(3) taking advantage of variable K-value strategy. Constant K is not suitable for all traffic patterns, which causing a greater probability of miscarriage of classifying and increasing errors. Different traffic patterns which is also different clusters can reduce the prediction error with different values of K.(4) adding a feedback mechanism. Adding error into the factors of distance metric guidelines can regulate non-parametric prediction model in order to reduce error.Finally, the algorithm is simulated. Simulation results show that the improved algorithm is obviously superior to the original algorithm.