This guest post is from Data61’s Advanced Data Analytics in Transport (ADAIT) team.

 


 

Transport is a trending topic, with Lyft and Uber looking to file for ‘initial public offerings’ (IPOs) shortly and companies putting out new products every other day to address various problems in transport – and it is about to get even hotter. CSIRO’s Data61 has been working with the New South Wales government for some time, using AI and sophisticated simulation modelling techniques to reduce congestion and help better manage responses to incidents which affect network flow.

Dr. David Harris, Research Director and Dr. Chen Cai, Research Group Leader, were recently invited to address the Standing Committee on Infrastructure and Cities, discussing the future of mobility from perspectives of alternative energy and AI in congestion management, providing insights into Australia’s future, as well as emerging trends in these areas around the world.

One of the topics Dr. Cai discussed was the ongoing joint program of work between Data61 and Transport for NSW, in which the Advanced Data Analytics in Transport (ADAIT) team is developing an innovative AI engine as part of the Intelligent Congestion Management Program (ICMP) initiated by the NSW government. Along this journey, the team has had to build up an ever-increasing set of capabilities and pool of knowledge, to allow them to deliver a unified predictive analytics and decision support platform.

This is the scenario we set out as a target: predict for 30 minutes and act in five minutes

Dr Chen Cai

One of the important problems to solve has been travel speed prediction.

Traditionally tackled by controllers using individual data sources, such as CCTV cameras and road sensor data, fed into regression or time-series forecasting models. These techniques fail to leverage the large volume and rich variety of transport data that can be analysed using modern data, engineering, and machine learning techniques.

By ingesting and fusing together large amounts of disparate data, the ADAIT team has been able to apply cutting-edge deep learning to achieve rapid, high-performance speed predictions for the road network, under normal operating conditions.

15 minute speed prediction on Victoria Road

15 minute speed prediction on Victoria Road

Often, the most interesting conditions to study are when the road network is not operating under normal conditions.

If there is a special event, road works or a traffic incident. Traditionally AI models struggle with infrequent non-recurrent events such as these, due to a lack of training data. The ADAIT team has developed multiple approaches for producing quality predictions under those conditions, including integrating traditional traffic simulation to analyse critical non-recurrent events. The simulation can run multiple scenarios, utilising preconfigured response plans, and compare the outcomes for travellers.

Incident detection and impact analysis framework.

Incident detection and impact analysis framework.

Traffic simulations are time-consuming to calibrate and execute, and hence should not be the only tool for dealing with non-recurrent events. One of the ADAIT team’s recent research focus has been to create more effective machine learning models for data-driven approaches.

The team has investigated different configurations of machine learning models, such as including higher order features and hidden nodes, which may provide a better representation of the data, and allow distinct non-recurrent events to be abstracted for use in more generic situations.

This should increase the accuracy of the machine learning models and provide better data-driven traffic network predictions, under more diverse operating conditions.

“We are in an unprecedented era and we have a lot more detail than before”

Dr Chen Cai

Data engineering is another significant part of the ADAIT team’s focus. They are partnering with leading data providers, to leverage a large number of disparate data sources, and are currently processing many millions of records each day. They use a microservices architecture, with Apache Kafka at its core, as the base platform for all of their data science and predictive analytics applications.

By having a strong cross-functional group of researchers, data scientists and engineers, the ADAIT team have been able to make inroads into furthering cutting-edge research in the AI and Transport space, while also being able to immediately reap the rewards, through practical application of their work to numerous industrial transportation projects.