Predicting 60 Minutes Ahead – A Case Study in Predicting Traffic Speeds and Volumes

Swinburne University of Technology’s Predictive Traffic Intelligence Study used traffic speed and volume data from Intelematics traffic data to demonstrate traffic prediction up to 60 minutes into the future. 

Hussein Dia, Professor of Future Urban Mobility, and PhD student Rusul Abduljabbar performed a multi-horizon traffic prediction study sampling Intelematics traffic data from Hoddle Street, Melbourne, extracting:  

• • Historical traffic data in 15-minute time segments; 

• • Over the 12-month June 2020 to May 2021 period; 

• • Looking at both Northbound and Southbound traffic; and 

• • 70,072 data observations. 

Swinburne’s traffic prediction model was developed as part of Rusul Abduljabbar’s PhD studies in Smart Urban Mobility/Artificial Intelligence (AI) in Transport. The model demonstrates the feasibility of using the pre-configured Intelematics dataset to forecast traffic conditions up to 60 minutes into the future, with high-resolution accuracy for predicting speeds and volumes during free flow and high congestion times. 

Professor Dia and Ms Abduljabbar met with industry and government and identified that there was a capability gap in the traffic data market and that local councils, government agencies, transportation engineering firms all lacked the ability to predict traffic accurately.

As Professor Dia explains, “We chose one hour as the time horizon based on government feedback. They didn’t want to know about tomorrow’s traffic, they wanted to know about congestion in the next hour and what they could do about it. Some government agencies can see what is happening currently on their roads, but what they needed was a tool to see up to 60 minutes into the future.”

The pair set about developing a predictive traffic model that leveraged machine learning to predict traffic conditions by speed and volume. The first task was to find a large and representative dataset.

Enter Intelematics traffic data.  

Intelematics traffic data covers more than 36,000kms of NSW and Victorian roads. With over 2 trillion data points, Intelematics traffic data gives a top-down view of the totality of traffic trends, going beyond broad year on year changes or incomplete and costly surveys focusing on only one road at a point in time.

Intelematics traffic data is collected from various sources, including sensors, cameras and in-vehicle trackers, to produce the most accurate traffic data available. The data is cleansed and validated using Machine Learning algorithms to analyse traffic trends and patterns down to 15-minute time segments on major and minor roads. 

Ms Abduljabbar chose Hoddle Street as the study area (Image 1), as it is one of Melbourne’s busiest arterial roads and regularly experiences high rates of congestion, especially after lockdowns, with average daily speeds of 41 kph slowing to a low of 30 kph at 3.30 pm when there is on average 736 vehicles passing through the Johnston Street to Gipps Street section during that 15-minute window.  

Image 1 – Study Location Hoddle Street Northbound and Southbound

Ms Abduljabbar took Intelematics’ historical traffic dataset and built a Bidirectional Long Short-Term Memory Model and trained it over 12 months, splitting the data into the training dataset (taking the first 60%), and the testing and validation dataset (remaining 40%), and then tracking the predicted speeds and volumes against the target.

As Professor Dia outlines, “The model is a deep learning neural network that processes in forward and reverse time in order to predict. In this way the model trains itself to mimic human decision making and increases in accuracy the more you train it on data that represents field conditions, becoming better at pattern recognition and creating millions of data connections. The key in achieving accuracy is in obtaining a large and representative dataset, and Intelematics traffic data was ideal for this.”

Image 2 – Swinburne’s Deep Learning Traffic Prediction Model  

96% Accuracy Achieved for Predicting Traffic Volume

Swinburne’s predictive model achieved high prediction exceeding 99% for speed and above 96% for traffic volumes, based on the extracted data. These results were achieved for both northbound and southbound traffic.

A major achievement was that the predictive traffic model maintained very high accuracy for predicting speeds 60 minutes ahead. This is difficult to do, as the further out the time horizon, the harder it is to maintain accuracy. This result sets Ms Abduljabbar’s model apart from other transport models, such as predictive parking, where the availability accuracy dips after 20 minutes.

Focussing on the southbound traffic, the speed prediction rates demonstrate accuracy of over 99.8% for the time horizons between 15 minutes to 60 minutes. Image 4 below shows how closely the prediction (orange) matched the real-world target (blue).

Image 3 – Southbound Speed Prediction Accuracy Results

Image 4 – Southbound Speed Prediction Horizon 15 to 30 Minutes Into The Future

In terms of predicting the number of vehicles travelling south along Hoddle Street, Image 5 demonstrates the model achieved 98.93% accuracy for the 15-minute time window into the future and was 96.13% accurate when predicting 60 minutes out. Image 6 shows how closely the prediction matched the target.

Image 5 – Accuracy of Volume Prediction Results for Southbound Traffic Every 15 Minutes

Image 6 – Southbound Volume Prediction Time Horizon 45 to 60 Minutes Into The Future

Swinburne’s Predictive Traffic Intelligence Study demonstrates the feasibility of using Intelematics traffic data for accurate traffic modelling and predictions. “We now have a trusted traffic dataset that can be used in a practical context,” Professor Dia said.  

Predictive traffic and congestion is key for calculating the cost of delays for cities and local authorities, for retail precincts and transport agencies, who will now be able to predict the future cost of congestion and delays.  

The return on investment of using predictive models include: 

• • Total hours lost from delays 

• • Increase productivity and economic activity due to the reduction of wasted time spent in traffic 

• • Total amount of CO2 Greenhouse Gas emissions created during delays that could have been avoided 

Importantly, expected delays on major and minor roads can be determined ahead of time and reduce costs on lost time spent in traffic delays. When the total number of hours on the road, driving below the average speed increases, efficiency decreases.

This type of forward planning is key for businesses, officials and planners wanting to develop an evidence-based business case and funding bid that accurately projects the expected benefits to drivers, the environment, and the local economy.

• • Huge time-saving. No more taking one month to cleanse two months’ worth of open-source data and having to check 55 detectors each day to delete errors from the dataset. 

• • Lowered project costs. As the data is already available in a ready to use, easily accessible format, resources and spending are significantly reduced. 

• • Quality Assurance. Make better decisions with high-quality results that use comprehensive, multi-source data points, including open source, private and commercial probe data, cleansed with smart machine learning algorithms. 

For more information please get in touch with our team.