Case Study: Data-driven approach for the implementation of Variable Speed Limit

Researchers at the Institute for Transportation, Iowa State University designed a data-driven approach for designing the advisory speed messaging. To test the logic a commuter corridor between Huxley and Ankeny, Iowa with a posted speed limit of 70mi/h was chosen. The significant contribution of this research is an effective advisory speed logic for the corridor which can appropriately advise speed with respect to the present traffic and weather conditions.

The dynamic advisory messaging provided the recommended speed that is changed across time intervals. As time progresses many events occur such a severe snowstorm or accidents which can affect the imminent traffic flow. To ensure a smooth progression of traffic, the posted speed limit is varied as per the existing conditions. The implementation of advisory speed messaging attempts to help in reducing accidents, reducing speed differences, prevent congestion and providing a calm driving experience.

This method is highly data-driven with consideration of current traffic flow and weather conditions for designing an advisory speed messaging. Learning from real-time data allows the logic to rightly capture the abrupt changes in traffic speed. The design of this system is handled as a two-part machine learning problem. The first part makes use of the abundant data available to cluster the speed data into groups capturing the current scenario. The second part utilizes expert judgment to label the data with an appropriate advisory speed.

Iowa DOT manages over 500 temporary and permanent Wavetronix sensors spread across the state. These sensors provide good quality traffic data via XML feeds refreshed every 20 seconds. This high-resolution speed and occupancy data collected are used for the problem. This rich real-world data in conjunction with expert judgment helped this technique to perform better than the conventional methods.

Wavelet transformation was applied to remove any intrinsic noise present in the sensor data. The data has properties that change with time, which are preserved by the application of wavelet transformation. This smoothing effect depended on two parameters, namely the size of the sliding window and the scaling factor for speed and occupancy. When implementing this technique in real-time, a sliding window using a subset of historical data is used for smoothing data. There are two important parameters that affect the smoothing of data namely the size of the sliding window and the scaling factor for speed and occupancy. While testing the model, two sizes were used for the sliding window – 10 minutes of historical data and 20 minutes of historical data. The scaling factor was varied from 0.5 to 6 in increments of 0.5.

Supervised and Unsupervised machine learning methods are used in designing this logic. The logic takes to input the smooth speed and occupancy data from sensors, cluster them into groups and assigns an advisory speed limit based on expert judgment. For the first part of the problem, the data is unlabeled hence an unsupervised learning method – K means clustering was used. The aim of clustering is to group the data based on speed conditions. The choice of K (the number of groups discovered by clustering algorithms) is determined by minimizing the objective function (Equation 1) which depends on the smoothing effects and the display performance.

J = Speed smoothing effects + display performance = f (N, k_speed, k_occupancy, K)

(Equation 1)

Conducting a grid search over the parameters the minimized objective function was obtained at N=64, k_speed = 5.5, k_occupancy = 2 and K=4. The four groups identified were in alliance with DOT recommendations for four advisory speed limits: 70mi/h, 55mi/h, 45mi/h, 35mi/h.

Once the clusters of data have been identified, a labeled dataset is prepared where each cluster is labeled with an advisory speed limit by an expert in the field. With this labeled data, a supervised learning technique is employed to predict the speed for traffic conditions in real-time. Support Vector Machine algorithm was used to define a function that will take input the smoothed speed and occupancy data and provide an advised speed limit label based on cluster analysis and expert judgment conducted earlier.

Evaluating this machine learning technique, it was compared with the performance of an existing logic based on the following three criteria: Number of total messages, Short-duration messages, Fluctuating clusters. This data-driven machine learning approach performed better specifically in detecting a snow event and changing the speed limit to a more advisable limit (35mi/h) than the existing logic (55mi/h or 45 mi/h).

The real-time implementation of this solution is a Java-based program with a Weka data mining package for classification (Witten et, al., 2016), JWave (Scheiblich, 2017) for wavelet data smoothing, and TransSuite® ATMS web service to post messages to field dynamic message signs (DMS). The Iowa DOT assigned four dummy DMS locations for testing, so the proposed control was still working in “ghost” mode until approved.