Design Of Short-term Traffic Volume Forecasting System Based On Non-parametric Regression

Short-term traffic volume forecasting system is one of key subsystems of the Intelligent Transport Systems (ITS). Perfect performance of forecasting and meeting real-time requirement concern the effective realizations of traffic control and transportation induction system.The implementation process of the short-term traffic flow forecasting system which was based on non-parametric regression model was described in detailed in this dissertation. The main purpose of this dissertation is to enhance the non-parametric regression (NPR) for use in practice. This dissertation, therefore, provides detailed analysis in functional requirements and performance requirements of the forecasting system, the detailed data flow which was based on the collaborative process between different function modules, system architecture and static class diagram of some key function modules.Non-parametric regression is a forecasting technique similar to case-based reasoning that does not make any rigid assumptions about the data. In short, the method searches a collection of historical observations for records similar to the current conditions and uses these to estimate the future state of the system. Unfortunately, NPR has many short comings restricting its applications which focus on: ill-suited database structure, low searching efficiency, open loop structure etc. This dissertation improves the method to meet the requirements of practical application.The main improvements include:1. Using modified KD-Tree to store historical patterns improve the search efficiency, significantly reduced the time required for data search, improve the performance of the forecasting system.2. Taking into account the real-time system, while tradeoffs between accuracy and timeliness are subject to the specific application of the real-time system, it is essential for the forecasting system to have imprecise results instead of none at all. As a result, approximate nearest neighbor instead of nearest neighbor is used in this system. 3. Feedback loop based on four levels of error is used in this system. Depending on the level of error, system will adopt the corresponding measures. Forecast accuracy is improved as a result.4. The weakness of normal forecast generation is that it ignores all information provided by the distance metric. To address this concern, a weighting scheme has been used. In general the weight is proportional to the distance between the neighbor and the current condition.Forecasting system is not projected in a single section, but the entire transport network. A large number of parallel operations occur during operation of the system; also, some local changes of the transport network structures may occur in practice. So system uses a distributed architecture which finishes prediction task within different computers and doesn't change the system structures when changes happen.Finally, the author concludes the study of this dissertation and points out the direction of future study.