Estimation and Prediction of Origin-Destination Matrices for I-66

摘要

This project uses the Box-Jenkins time-series technique to model and forecast the traffic flows and then uses the flow forecasts to predict the origin-destination matrices. First, a detailed analysis was conducted to investigate the best data correction method. Four spatial correction procedures were examined for non-incident related detector data. The first approach, temporal correction, exploited the inherent temporal trend of historical traffic. The spatial correction based on linear regression (LR) - a proposed modification of a previous approach - uses the relationship between the individual detector flow and station flow. The third approach proposed in this study is also a spatial correction method. A unique feature of the proposed spatial correction procedure was incorporation of lane use percentage into the correction process through kernel regression (KR). As a comparison benchmark, the correction method based on lane distribution (LD) developed by previous researchers was included as the fourth method. To comprehensively compare the correction procedures, both systematic evaluation and random-error evaluation were conducted. After the results of systematic evaluation were analyzed, it was found that adaption was needed for the KR and LD approaches. Specifically, the individual lane flows provided by the detectors on particular general purpose lanes produced more accurate estimates. The two correction procedures (kernel regression and lane distribution) were revised in light of this finding and their station flow estimates were compared to those of the temporal correction and the LR approach at five error levels, which was considered as the random-error evaluation. After applying the temporal correction method to the data set, the station flow series was modeled using the Box-Jenkins time-series modeling framework. The station flow series was successfully modeled by an MA(2) model using the Box-Jenkins time series technique. It enjoys an MAE of 344.53 and MAPE of 7.07%. The limitation is that different models are needed for each station and the model may be different for a different time period. However, once the model is fitted, it can be used to forecast the traffic hourly flows, which were subsequently used to predict the origin-destination demands. The performance of QueensOD modeling is relatively good, which is evidenced by approximately 10% MAPE for a majority of links. Generally, the relative errors (MAPE) are around 10% with a few higher exceptions. This is possibly due to the fact that the QueensOD requires the actual travel time as inputs. The actual travel times were computed using the BPR function according to observed flows, which could be a source of error.