Machine learning based traffic demand estimation

• 主要作者: Wang, Xinxu
• 其他作者: Wang Dan Wei
• 格式: Thesis-Master by Coursework
• 語言: English
• 出版: Nanyang Technological University 2022
• 主題: Engineering::Computer science and engineering::Theory of computation::Analysis of algorithms and problem complexity
• 在線閱讀: https://hdl.handle.net/10356/160982

Traffic demand analysis is an important element in traffic planning and an important basis for solving traffic problems. With the development of intelligent transportation, traffic management is becoming increasingly informative, placing higher demands on further mining digital resources, realizing dynamic traffic demand analysis, and providing real-time information for traffic management. Traffic demand analysis mainly consists of dynamic origin-destination (OD) estimation and dynamic traffic assignment (DTA). OD estimation is essential for building micro simulation platforms with realistic traffic demand, with the aim of validating applications of intelligent transport systems, such as traffic signal control and urban planning. As input to Dynamic Traffic Assignment, a dynamic OD matrix generates time of day varying traffic demand. Dynamic traffic assignment captures traffic flow data for future road segments within the road network and provides the data basis for the development of time-varying traffic management schemes. This dissertation describes the methods for OD matrix estimation and dynamic traffic assignment. Firstly, the backpropagation method of the OD matrix and the big data-based acquisition method are introduced. Secondly, the dynamic traffic assignment method based on VISSIM simulation is introduced, and the travel cost model, the journey time model, the logit path selection model, and the impedance multivariate multipath assignment method are described. This dynamic traffic assignment simulation model can be dynamically assigned after inputting the OD matrix, and the traffic flow of each road section of the road network can be obtained by setting the detector. MATLAB was used to generate a large number of OD matrices that met the road conditions, and the machine model was trained using the OD matrices as the input and the traffic on each road segment as the output. The CNN model and LSTM model were used for training. The trained models are then used to make predictions to obtain the predicted traffic volumes for each road section. The accuracy of the machine model estimates is evaluated by comparing the predicted data with the data obtained from the VISSIM dynamic assignment. The comparison results show that the CNN model has an estimation accuracy of 82.03% and the LSTM model has an estimation accuracy of 83.24%. The LSTM model with better model performance was used as the traffic assignment system.