Traffic Speed Modelling And Prediction Based On Multi-scale And Multi-dimention Data

With the rapid development of economy of China, the issue of heavy traffic emerges, which brings about traffic congestion, traffic accident, air pollution and so on. To ease the overload of urban traffic, Intelligent Transportation Systems (ITS) provide a more safe and efficient transport network, via regulating the traffic flow. Among all technologies in ITS, the traffic speed prediction is the critical one, as it lays a solidary foundation for other applications, such as traffic signal control system, traffic guidance and so on. Recently, it is noticeable that an increasing number of taxicabs are equipped with in-vehicle GPS systems, providing 24-hour and wide-range information about location and speed to data-driven ITS. Since taxicab is one of the common travel methods for urban residents, reflecting dynamics of urban traffic when driving through every corner of city, its GPS data has drawn a great deal of attention from researchers. Specifically, by analyzing and mining the mass of GPS data, they compute vehicle speed and predict the future traffic speed. However, to the best of our knowledge, the existed prediction methods fail to accommodate the long-term prediction, especially when occasional events, say the bad weather conditions, occur. To tackle that problem, we leverage over 88000 GPS data of taxicabs in Hangzhou, as well as the data about weather conditions and special day, and then propose a hybrid traffic speed modelling and prediction framework.To begin with, we establish the platform in order to efficiently pre-process multi-dimension data, which is on the basis of Hadoop distributed computing system. Hence, given the sparse GPS records on a certain road, we divide the traffic network into numerous segments and average the speed data within one hour in one segment, so as to conduct the correlation analysis based on these average speeds. Thereby, we find out the correlations among multi-time-scale traffic speed data, and the profound effects caused by occasional events such as weather condition and special day. Subsequently, based on the correlation analysis and inspired by the methodology of system identification, we propose a hybrid traffic speed modeling and prediction framework which contains a systematic recursive model identification algorithm to derive the model order and parameters, and takes multi-time-scale historical traffic speed data and multi-dimension random factors as inputs. After all, we set up the traffic prediction models for major road segments in Hangzhou respectively, and evaluate the prediction performance based on the dataset under various settings. The extensive results indicate that our framework outperforms with regard to the long-term prediction. Besides, compared with the benchmark method Supportive Vector Regression (SVR), our method performs outstanding.