Forecasting Traffic in Athens

Work in progress

Mon Nov 24 12:17:57 UTC 2014

Over a 390 day period between 2013-10-08 and 2014-11-03, I recorded average vehicle speed, sampled every 5 minutes from a network of 45 loop detector sensors in Athens. The data were not recorded directly (e.g., from ATMS (Athens Traffic Management System)), but instead, indirectly from a traffic information service provided by Naftemporiki newspaper.

Some time ago, I used a much smaller subset of this data (based on a few days) to generate a crude traffic forecasting model for use in my taxi cost estimation tool which I describe in rough detail on my old blog, here. I later used the same model for the taxi component of my public transport exploration tool zee.gr, described here. The point of this project is to investigate some alternative models, and learn some new things along the way. Using the harvested data, I intend to:

• Take a look at different mapping tools besides google maps (my previous projects have been pretty much biased to google maps api.)
• Do some exploratory data analysis in R to discover the time lagged relationships between each sensor.
• Have a look at spatial interpolation: The data are from specific points, however it would be interesting to interpolate these readings to roads that do not have sensors. IDW (Inverse Distance Weighting) and Kriging look promising.
• Develop a traffic forecasting model based on the spatial temporal relationships of all sensors to predict average traffic speed at increasing time horizons.
• Maybe investigate the impact of various events, such as public transport strikes on traffic flow.

This is an ongoing (mainly for fun) project. Any code developed will be hosted in this github repository: github.com/phil8192/athens-traffic. So far the repository just contains data harvesting code, some pre-processing and some initial cross section summaries.

Interesting papers

• Nice meta review: E.I. Vlahogianni et al., Short-term traffic forecasting: Where we are and where we’re going. Transportation Research Part C 43 (2014) 3–19.
• Neural network specific: E.I. Vlahogianni et al. Optimized and meta-optimized neural networks for short-term traffic flow prediction: A genetic approach. Transportation Research Part C 13 (2005) 211–234
• Athens specific, nice paper on statistical approaches: G. Kamarianakis, P. Prastacos., Forecasting traffic flow conditions in an urban network: Comparison of multivariate and univariate approaches. 82nd Transportation Research Board Annual Convention (Paper Number 03-4318). 2003.
• W Min, L Wynter, Y Amemiya., Road traffic prediction with spatio-temporal correlations. IBM Research report. Mathematics, 2007.

Harvested Data

The dataset consists of 390 days of 112,608 5 minute observations for 45 individual sensors, totaling 5,067,360 data points. Each observation is the average speed (determined by the sensor source) in kmph for a road section. Of all of the recordings, the following histogram shows the average speed in Athens. The mean speed is 51.19 Kmph.

Sensors

The actual sensors are on arterial roads in and around the centre of Athens. From what I have read so far in various papers, this is a common configuration; Side roads and suburban areas are neglected. This makes the problem of spatial interpolation somewhat more difficult. The following map quickly shows the (intersecting) road segments included in the dataset.

Hierarchical Clustering

Given the average weekday speeds for each of the 45 sensors , the following graph shows the individual sensors arranged into hierarchical clusters. The congestion at 9am is shared on all roads. After 9am, roads in the centre remain congested, while others clear. Note that Kifisou is the main road running from Metamorphosi towards Peiraias.

Cross sections

So far, I have only looked at cross sections of the data. To begin with, I have separated weekend data from weekday data, and further still into Saturday and Sunday. The following 3 graphs show the average vehicle speed (white line) for weekday and weekend traffic. For example, the 12pm point shows the average vehicle speed at 12pm for an entire year in the weekday or weekend. The red line is a loose approximation/fit to these observations. During the week, the traffic is worst at 9am, after which it improves until 11am, after which it remains more or less constant until 6pm, after which it begins to improve. On Saturday, traffic is worse at around noon. And on Sunday, the average vehicle speed dips again at noon, but then again at 8pm.