Data and simulation: Improving Perth’s road network performance
iMOVE’s Improved network performance prediction through data-driven analytics and simulation project began in July 2017 and was completed in March 2020. The project was led by Main Roads Western Australia, with research provided by the University of Western Australia’s Planning and Transport Research Centre (PATREC). The final reports generated by the project are downloadable further down this page.
Main objectives
The project had three main objectives:
- Develop mathematical and empirical models to predict short-term traffic flow characteristics for individual road sections.
- Investigate the probability of applying perimeter control (or gating) strategies based on Macroscopic Fundamental Diagrams (MFDs) to guide network operations at a sub-regional level.
- Develop simulation models to identify possible impacts to the operation and planning of Perth’s freeways due to the introduction of Autonomous Vehicles (AVs) and Connected and Autonomous Vehicles (CAVs).
Machine learning and congestion prediction
This sub-project used machine learning to predict average speed for individual road sections in the Perth metropolitan road network up to a horizon of 75 minutes in advance. This capability will enable stopping congestion before it occurs and faster incident detection and recovery. Looking ahead, the aim is the ability to now-cast, meaning ‘the prediction of the present, the very near future and the very recent past’. This is needed because of the delay in data acquisition and cleaning; and the dynamic nature of network operations makes predicting the very near future valuable. Knowing what is likely to happen could help traffic operators making more evidence-based decisions.
As a result of the work on this sub-project, it was proven that the developed machine learning technique performed well against the benchmarks. The highlight finding was that accuracy does not decrease dramatically with an increasing prediction time horizon, e.g. during the AM and PM peaks, predicting 15 minutes ahead will produce an average percentage error about 9% while for 75 minutes it is just above 10%.
By contrast, the performance gap with benchmarks becomes more pronounced with increasing prediction timespan. In this instance the models worked using offline historical data. A move to use real-time data in the future is desirable, but requires overcoming the current obstacles of sufficient funding for livestream data acquisition and necessary data infrastructure.
The models could also be further improved by:
- Inclusion of more data, particularly an entire year of seasonality
- Improving data quality, such as further calibrating VDS detectors
- Fusing multiple speed data sources (using the model speed fusion approach developed in a sister project)
- Development of a deep learning model
Perimeter control
One of the cornerstones of smart freeways is the use of ramp metering to regulate traffic inflow at on-ramps. This prevents flow breakdowns on the freeway so it can remain reliable while delivering higher throughput. This iMOVE subproject investigated the possibility of applying perimeter controls, a similar concept but expanded to the whole network, including arterial roads.. What’s perimeter control?
Perimeter control (also known as gating) works by dividing the network into multiple zones and regulating their flow exchange at the boundaries. It aims at load-balancing between zones across the network to achieve a stable and optimum operation at the global level. Controllers prevent overflow of traffic into busy zones by leveraging spare capacities in less busy zones as temporary storage space. This contrasts with local congestion relief strategies that focus on individual pinch points, which can result in pushing too much traffic downstream and creating another bottleneck.
The effective implementation of perimeter control requires a good understanding of the traffic behaviour of each zone. Macroscopic Fundamental Diagrams (MFDs) are commonly used for such purposes. They describe underlying relationships between a zone’s speed, flow and density at the aggregate level. Knowing the characteristics of MFDs could help traffic engineers gauge the current traffic state and prevent it from going beyond capacity that will result in traffic congestion. The accurate measurement of MFDs has become possible in recent years with the advent of ‘big data’
For perimeter control to really work, it needs to be applied at the right zonal level. For Main Roads’ current operational purposes, Perth’s metropolitan road network is divided into four performance sub areas which are too large for MFD-based perimeter control since they very rarely reach critical density as a whole. These four large zones were bisected until desired MFDs with low scatter and clearly defined critical density were achieved, which occurred at 38 zones.
Zones with high traffic demand tended to have more usable and well-behaved MFDs that clearly indicate a critical density or ‘tipping point’ while still featuring low scatter. Conversely, zones with low traffic demand tend to have less-than-ideal MFDs with higher scatter and no clear indication of critical density. They are less suitable for the implementation of perimeter control.
Simulations of perimeter control using mathematical models demonstrated that ‘…regulating traffic flows at boundaries by means of signalling, can avoid flow breakdown of the congested zones.’ Although the traffic was slowed in zones that act as the buffer, the whole network performed substantially better. In a 12-zone model, travel time decreased by 12%, though the caveat on this is that ‘… the numbers should not be taken literally since the models are hypothetical and include many assumptions. They are only intended to illustrate the potential benefit of gating.’
More research on this is recommended before making this change to the network’s roads operational, and its promise does make a solid case for additional work. The PATREC team is currently testing the idea using a Perth CBD traffic simulation model, which will produce more realistic results than the previously developed more abstract mathematical model.
If successful, it can maximise the productivity and reliability of the whole network by utilising spare capacity in zones with low demand-to-capacity ratio to alleviate stress for those under high demand. The productivity gain and avoidable cost of unnecessary road expansions could generate significant social, economic and environmental benefits.
Impact of AVs and CAVs
We still don’t know exactly when they will join Australian road networks, but Automated Vehicles (AVs) and Connected and Automated Vehicles (CAVs) will one day appear here. And their appearance is both an opportunity and challenge.
… any long-term investments without adequate consideration of their potential impact are inherently risky. It is reasonable to question whether planned major transport infrastructure will be appropriate in accommodating and facilitating a fully automated future.
This last sub-project ‘… marks the first step towards modelling the potential traffic impacts of AVs and CAVs on Perth’s freeways and arterial roads.’ Modelling was conducted on the Mitchell Freeway, Stirling Highway, and Canning Highway.
There are many uncertainties associated with AVs and CAVs so how they will perform is still largely unknown. Some even argued1 that riders could slow them down for more comfort if they want to work while their vehicles are driving themselves, which could cause worsening traffic operations. To address these uncertainties, the AV and CAV models were given a wide range of performance parameters. An optimisation algorithm was used to help quickly find the best- and worst-case scenarios.
The results show that despite being tested using a wider range of performance parameters than human‐driven vehicles (e.g. lower acceleration and deceleration rates to improve comfort), both AVs and CAVs still performed better at 100% market penetration rate than the current base case with 100% human-driven vehicles. And AVs and CAVs benefited freeways more arterial roads.
Meanwhile, CAVs perform significantly better than AVs in all situations, not only in terms of shorter delay time but also higher travel time reliability, leading to more reliable journeys. Instead of predicting an exact future, which is impossible, the report established the range of possibilities for planners and policy makers to consider.
In terms of what’s next, the report advises that ‘… the technology readiness for AVs/CAVs should be regularly assessed and business cases should consider them when mass deployment is within reach,’, and also that the report should be updated ‘… new data or information about AV and CAV driving behaviour is available,’
Comments were also made about the implementation cost and maintenance cost of Vehicle to Infrastructure (V2I) and Vehicle to Vehicle (V2V) technologies.
Comparing V2I, V2V as a consumer technology has the advantage of not requiring expensive public infrastructure and its associated maintained costs. Since our results show that V2V plus some simple V2I technologies alone could make significant improvement to the network, further research needs to examine whether this diminishes the necessity of more sophisticated and expensive V2I technologies.
Lastly, another idea targeted for further investigation is the research being extended to investigate dedicated AV lanes on Perth freeways.
Footnote
- Le Vine, S, Zolfaghari, A, & Polak, J 2015, ‘Autonomous cars: The tension between occupant experience and intersection capacity’, Transportation Research Part C Emerging Technologies, 52, pp. 1–14.
Download the reports
For your copy of the three final reports generated by this project please click the buttons below.
.
Discover more from iMOVE Australia Cooperative Research Centre | Transport R&D
Subscribe to get the latest posts sent to your email.