Traffic Performance Evaluation Using Statistical and Machine Learning Methods

Abstract

Fast and safe movement of people and goods is one of the key objectives of every efficient transportation system. The rapid expansion of urbanization, economic growth, and the increase in car ownership per capita, significantly impacted the efficiency and safety of the transportation network. In the US, a growing portion of all daily travel is happening in congested or oversaturated road conditions. Congestion, excessive queue, and spill-back will negatively impact environmental pollution and could lead to network-wide traffic system disruption. Therefore, transportation agencies and authorities are actively investigating innovative regulations and technologies to tackle the growing issue of traffic congestion. However, prior to full adaptation and implementation of any new technologies or regulations, policymakers and transportation engineers are required to collect reliable information regarding their effectiveness and conduct cost-benefit analyses.A common approach for conducting cost-benefit analyses for the adaptation of any new technologies or regulations is a before and after study. Before and after studies have been commonly used in many real-world evaluations, including traffic and transportation evaluation projects. Determining sufficient data samples, eliminating the influence of interfering factors, and proactively evaluating traffic performance are important to draw accurate conclusions and provide insights for decision-making in traffic performance evaluation. However, very little research has been carried out on developing a concrete methodology that can be used to systemically evaluate traffic performance in before and after studies. This dissertation aims to establish a general framework that employs statistical and machine learning methods to improve the accuracy and effectiveness of traffic performance evaluation through before and after studies. This dissertation consists of the following components. One of the major challenges faced when conducting a reliable and robust traffic performance evaluation is data sample size determination. Therefore, the first component of this dissertation aims to determine the data sample size needed for accurate traffic performance evaluation. Data requirements include the type of data and the duration of the data collection period that is needed for assessing traffic performance evaluation. None of the existing research statistically evaluated the data requirements for traffic performance evaluation. In addition, no solid approach exists for determining the data collection sufficiency for a robust traffic performance evaluation. In this component, a non-parametric statistical approach is proposed to determine the amount of data needed for traffic performance evaluation. The proposed approach is robust to the underlying distribution of the random variable. That is, the accuracy of the model is insensitive to the data distribution. For validation purposes, three active ramps along State Route 51 in the Phoenix metropolitan area of Arizona were selected as the case study. Based on the analysis results, it was found that data be collected for two consecutive months to be able to effectively evaluate the responsive ramp metering strategy. That is, the first month of data collection will be set as the accommodation time, during this time the frequent commuters on SR-51 will adapt to the new ramp metering strategy and the second month of data collection will be set as the minimum amount of data required to assess the change in the ramp metering strategy. While assessing traffic performance using before and after studies, usually the outcome is directly or indirectly impacted by factors such as seasonal factors, holidays, and lane closures. These factors might interfere with the evaluation process by inducing variation in traffic volumes during the before and after periods. In practice, limited effort has been made to eliminate the effects of these factors. The second component provides an innovative solution to eliminate the impact of interfering factors while conducting before and after evaluation studies. In the before and after study, the assignment of samples is non-random across the treatment and control groups. Therefore, when estimating the effects of a treatment on an outcome, the result might be biased. One potential solution to resolve the biased caused by non-random assignments of treatment and control groups is using Propensity Score Matching (PSM). This component aims to eliminate the influence of interfering factors in traffic performance evaluation by proposing a PSM-based method that uses extreme gradient boosting (XGBoost) to estimate propensity scores. The proposed methodology applies a t-test and Kolmogorov–Smirnov (KS) test to conduct a balancing test for determining the sample size for control groups. Besides, statistical analysis results from the t-test and KS test are employed to choose the most accurate one when comparing different PSM methods. The proposed method was applied to a corridor in the City of Chandler, Arizona, where an advanced traffic signal controller has been recently implemented. The results indicated that the proposed XGBoost-based PSM method was able to effectively eliminate the variation in traffic volume caused by the COVID-19 global Pandemic during the evaluation process. In addition, the results of the t-test and KS test demonstrated that the proposed method outperforms other conventional propensity score matching methods. In some scenarios implementing technologies and before and after studies are costly; therefore, transportation agencies and authorities need to conduct proactive traffic performance evaluations. In the third component of this dissertation, a novel proactive traffic performance evaluation approach is proposed. Proactive traffic performance evaluation helps transportation agencies evaluate the effectiveness of their new technologies before field implementation. Simulation-based techniques and traffic forecasting methods are two of the commonly used methods for predicting traffic states. Simulation-based techniques can predict traffic states by modeling traffic conditions in a simulated environment. However, the model calibration process is time-consuming and labor-intensive. Traditional traffic forecasting methods are incapable of predicting traffic states when the transportation network infrastructure is influenced by new technology. In addition, the use of traffic forecasting methods in evaluating the effectiveness of ITS technology, such as ramp metering, is somewhat limited in the current literature. To address the above-mentioned research challenges, this research component introduces an innovative approach to proactively evaluate ramp metering performance. Ridge regression and Two-stage TrAdaBoost.R2 algorithms are adopted in this approach to obtain the research objectives. The effectiveness of the proposed approach was examined by comparing the results with real-world traffic data collected from the southbound of State Route 51 (SR51 SB) of the Phoenix metropolitan area. The proposed approach exhibits superior spatial-temporal transferability for new freeway segments in comparison to traditional machine-learning methods. In summary, the proposed statistical and machine learning methods in this dissertation could help transportation engineers and policymakers conduct accurate traffic performance evaluations. Furthermore, these statistical and machine learning methods provide insights into accurately measuring the overall mobility of transportation systems, addressing the limitations of current traffic management strategies, and effectively limiting these widespread losses brought by inefficient traffic systems.