Although both the background terminology and the technology are far from ‘new’, advances in Artificial Intelligence (AI) are rapidly changing the world as we know it – much of this change is taking place without us even noticing.
With the increasing power and sophistication of computing and sensor technology, the accumulated data we have about our everyday movements, and those of our assets, is mind-boggling. Terabytes of data are created every day, but the question is how to make sense of it all and turn it into useful and actionable intelligence. This question is something not all public transport authorities and operators have fully got to grips with.
The AI train has gained rapid momentum in recent years, both in terms of business usage and results. We’re now seeing the technology being used for disease detection, climate change analysis, air traffic control, social media, and retail marketing. In a 2021 report by digital training and consultancy specialist O’Reilly, 26% of businesses surveyed say they have reached ‘mature’ stage, incorporating AI and Big Data into their products and processes.
In early 2020, UITP quizzed public transport authorities about their usage of the technology. Around a quarter of respondents confirmed they were either using, testing or offering AI in certain areas – whether this be real-time operations management, scheduling, journey planning, customer analytics, fraud detection or safety management.
While 25% seems like a high number, transport trails other sectors, with recent studies suggesting that retail comes out top when it comes to AI maturity (40%), closely followed by financial services (38%) and telecommunications (37%). Given the opportunities, public transport appears to be lagging behind, especially as it is currently only being explored by the larger authorities and operators.
Johan C. Haveland, former Director of Passenger Transport at Bergen’s Bybanen light rail system, is at the cutting edge of harnessing computing power and advanced algorithms for public transport applications. After leaving Bybanen in 2020, he founded Asistobe to specialise in the development of tools to optimise transport planning and scheduling. Speaking at this year’s UK Light Rail Conference, Mr Haveland outlined an exciting future for public transport with AI – and the very real benefits it could deliver in the surprisingly near future.
Mr Haveland explained: “This could be one of the critical solutions that efficiently unlocks the value of data to improve the quality and efficiency of the sector”.
Making the most of Bergen
At the root of all this is data, and the use of applied mathematics and computer science
to analyse it to make data-driven decisions. Mr Haveland has an abundance of experience in this area, having utilised such methodologies to help Bybanen achieve remarkable efficiencies.
The June 2010 opening of the first section of the Bybanen light rail system was relatively simple – a single 9.8km (6.1-mile) north-south line from Byparken terminus in the city centre to the town of Nesttun. Automatic passenger counting was featured on just two of the
28 Stadler Variobahn trams at the launch. The system has since grown in significant increments, firstly with a 3.6km (2.2-mile) five-stop extension to Lagunen in June 2013, followed by a further 7km (4.3 miles) to reach Bergen Airport in April 2017.
Services were crowded in the city centre from the start, but towards the airport the traffic thinned out considerably. Trams ran at a five-minute frequency from end to end, although with so much ‘empty running’, the city sought more effective and efficient solutions as the network evolved.
Asistobe’s task was to make the best use of the network’s assets and maximise their cost-effectiveness. “We began by making graphs to show passenger distribution,” Mr Haveland explained. “I did this with pen and paper initially, running different scenarios and different line structures on the single-line system to show how we could save money and move more passengers at the same time. We then got Big Data analysis to prove I was right.”
With the first three projects open, stage four at the construction stage, and planning for stage five underway, Asistobe was asked how to best plan the new routes, devise operational models such as where to turn the trams, where bus terminals should be in this whole infrastructure, and to develop a wider transport network plan. It quickly became clear that pen and paper wouldn’t be the best way to do this.
“We did corridor analysis of how people actually travel in Bergen,” Mr Haveland described, “using algorithms to see where people are travelling, and how the routes operate. This revealed a lot.
“We were able to reduce the number of kilometres driven, from 3.75m km per year to 2.74m km per year. This provided an obvious reduction of 23.3% in operational expenses, and we were able to transport more people at the same time. This was not achieved by over-crowding the trams, it was all about smart scheduling and making better use of
The structure of public transport in Scandinavia follows a common European model where tram, bus and ferry networks are run by the same entity. Efficiencies and cost-savings are therefore considered across a city-wide or region-wide network. In Bergen the transport authorities wanted to know where they could invest in light rail, but also where cost-savings could be made on bus routes to help allow this to happen. The question was how much money could be saved.
It is important to emphasise that this exercise was not for the sake of saving money alone, but because the need to invest more wisely was the top priority.
“You need advanced algorithms to solve such equations,” Mr Haveland argued. “No-one can do this with pen and paper or spreadsheets. It’s like an advanced game of chess, where we’re predicting the best move at any given time, except with millions of passengers, thousands of pieces (trams and buses), and a playing field with ten thousand stations.”
It’s all about time
The next step was to look at another aspect where AI can make a very real difference: real-time operations management.
The initial strategy in Bergen when an incident occurred – for example, a tram breaking down in a tunnel – was to shut the entire network. Time would be taken to assess the available resources and evaluate possible solutions. After a discussion, and around ten minutes of elapsed time, those decisions would then be communicated to the drivers and other personnel out on the network. After another 15 minutes, a new intermediary model would be put in place; operation would slowly get back to normal thereafter.
The value in employing AI to strategise incidents, rapidly and dispassionately, dramatically reduces the time taken to assess such situations and make those decisions. Automated systems that have played out such eventualities thousands of times mean it is now possible to have a ‘Plan B’ up and running a lot more quickly.
“The OCC [operations control centre] obviously makes a lot of decisions every day,” Mr Haveland described, “with decisions based upon past experience such as time of day, fleet position, expected passenger flow etc. We did a lot of mapping of this data and created basic algorithms to help make these decisions faster, because in a complex environment such as a control centre it’s very hard to make those necessary decisions quickly under pressure as human beings.”
With the algorithms in place, AI was able to make those resolutions instantly and following clear established logic. With the time to make a decision taken down to zero, this speeded up the communication with the drivers to put a new temporary operation in place: “The total time to have the network running again has been cut to three minutes. The reduction in time taken to make rapid decisions throughout the day immediately following an incident also has the knock-on effect of cutting the time taken to resume normal operation.”
Although the OCC algorithms aren’t used in a live environment, yet, the proof of concept is there and Asistobe is in discussions with a number of operators about pilot schemes.
Yet while AI is currently being used to plan routes, track alignments and operational models – plus assisting in making incident resolution decisions – its potential is far wider.
Next steps for AI
No assessment of the future role of technology in public transport would be complete without considering its use for autonomous operation.
Probably the most high-profile example was the Siemens demonstration of a prototype driverless tram in Potsdam during the InnoTrans railway trade fair in 2018. Since then the technology has been tested in ‘real road’ traffic conditions along a 6km (3.7-mile) section of the Potsdam tramway. Various elements of depot operations have also been moved to autonomous running in trial form and a number of similar pilot initiatives are running across Europe and Asia.
In August, the consortium of Siemens, operator ViP Verkehrsbetrieb Potsdam, Germany’s Institute for Climate Protection (IKEM), Energy and Mobility and software provider Codewerk announced that it is aiming for commercial viability of its autonomous depot management system in 2026. (See News for more detail).
Uwe Loeschmann, CEO of ViP, said: “Autonomous driving along the tram route and within the depot relieves our drivers and increases the safety of our passengers and other road users… Autonomous tram operation in our depot with the AStriD (Autonomous Tram in Depot) system opens up the possibility of automated cleaning, supply and parking processes with central control and increased operational safety.”
Believed to be of national importance, these studies are funded by Germany’s Ministry of Transport and Digital Infrastructure, with IKEM exploring the complex legal and economic issues related to the project.
These obstacles are the key inhibitors to further progress, Mr Haveland explained: “The driverless tram relies on so much technology that I don’t see these as the key to teaching trams how to drive on their own. The key is how to handle the algorithms and do the calculations in real-time to ensure safe operation. I still haven’t seen any tram operate fully autonomously, yet, there’s still a requirement for a driver next to the machine ready to brake.”
Echoing previous TAUT commentators, Mr Haveland believes that societal, ethical, legal and regulatory considerations are the key barriers to further automation. These are complex challenges. “Who is responsible if something happens?”, he asks, “The supplier of the tram, the OCC, the municipality?”
A driverless future for transport is very much being driven by the automotive industry as the complexity of routes, intersections, interactions with other traffic and the wider environment is much more ‘open’ with road vehicles, he suggests. So what can we learn from work already underway in other sectors?
Running an autonomous tram or light rail vehicle is clearly simpler than running an autonomous car given the fixed guideway infrastructure and fixed routes. “In reality,” Mr Haveland pointed out, “what we’re looking for is what is different. I think their algorithms will work for rail, but the problem will be finding places to demonstrate these systems, get enough kilometres driven, and resolve any technological challenges to prove it’s safe.”
This is where we come back to Bergen, as the city’s single line could offer a perfect testbed for such evaluations.
Looking back hundreds of years to the application of the very first machines, the argument has often been made that technology will displace the need for people, reducing employment opportunities. In public transport, the logic suggests that drivers and other operational staff will no longer be needed when driverless trams finally get their approval for commercial running.
However, Mr Haveland says that the response to the publicity raised when talking about the autonomous tram in Bergen has only ever been positive. “There would still be plenty of jobs in the system,” he said. “You will always need people. I got a few comments from the drivers, but they were friendly comments – maybe they thought it was so far in the future that it wouldn’t affect them.”
However, with advances made as each year passes, the impending arrival of autonomous vehicles is “not a question of if they will be driverless, it’s when,” he concludes.
A digital crystal ball
Previously, tramway maintenance has relied on a rigid schedule of periodic checks, where tasks are dictated by the vehicles’ usage and the mileage covered. With trams in Bergen taken in for checks every 20-25 000km (12 550-15 500 miles), more recently there has been a shift towards condition-based maintenance, utilising sensors on both the vehicles and the infrastructure to indicate what needs doing and when.
There are further efficiencies to be achieved as a lot is still done on a schedule basis, with Mr Haveland making the case that vehicles may still be taking trips to the workshop more regularly than they need to.
Developments in AI mean that a paradigm shift is fast approaching, he says, where predictive rather than reactive maintenance will soon become the norm. “Sensor technologies are moving so fast and even the most advanced systems are far more available and affordable than they used to be. There are many tools around to collect the data and handle the millions of data points every day. The systems can handle far more data than is currently being directed their way, too.
“Add that to machine learning and we now have the ability to predict when something will break, what other systems will be affected by that break, how long a malfunction will take to fix and what how much it will cost.”
With such a vast number of technical, operational and environmental scenarios, it is extremely difficult to predict what could possibly go wrong on a public transport system. This time-consuming process can also lead to overly risk averse planning, asking the ‘what if?’ questions for situations which may happen infrequently.
Each of the tens of thousands of components in a network can in turn be affected by temperature, vibration, pressure, humidity etc. To find the weak points, Mr Haveland and his partners at Norwegian start-up Nortech AI have built predictive systems for a large industrial company, installing sensors on equipment and creating algorithms to understand the problems working for and against it.
By leveraging machine learning, they can now predict when things may go wrong based upon any number of influencing factors. Establishing the baseline ‘normal’, the deployed algorithms were able to forecast that a piece of equipment would break down on a particular day in November… eight days before it actually did.
“In a normal situation,” Mr Haveland explains, “this equipment would just break down, but by then it’s already too late. If you’re able to predict that something will break in eight days’ time, you’re able to move it into your maintenance regime and have all the parts and components in place without disrupting the operation in any way. As we get more and more sensors, predictive maintenance will play a much bigger part in how we manage our transport networks.”
The 2017 report Artificial Intelligence: The Next Digital Frontier? from consultancy McKinsey explored in detail how businesses can save money and improve operations based on this maintenance model. The report claimed 10-15% savings on operational costs with a move to condition-based maintenance, and Mr Haveland estimates that by using the tools available today, a move to predictive maintenance could result in a saving of around 25% of operational cost.
More than just cost-cutting
As a business case for light rail, based purely on the suggested cost savings highlighted so far, shifting some operational elements to AI surely appears worthy of investigation.
With estimations of 23% savings via smart scheduling and operation – using 10-15% fewer trams is in itself a major saving – 25-35% via the use of autonomous tram and depot operations, and 25% from the maintenance budget, the overall sums saved would easily run into the millions.
While employing AI and Big Data analysis for real-time operations management doesn’t directly result in operational cost savings, it does impact on the quality of service, and ultimately, passenger satisfaction. In all, on a purely commercial cost comparison basis, operational savings could be reduced from 35% at the lower end, to 70% at the higher end, by moving all of these areas to an AI-based regime.
“The business case for transport authorities and operators is significant,” Mr Haveland concludes: “However, the money wouldn’t necessarily be saved. In Bergen those funds are then ring-fenced for reinvestment back into the public transport network”.
Enticingly, he goes on to pose one final question: “How much better would the public transport systems in our cities be if we could really reduce our operational expenses by
such high amounts?”
Article appeared originally in TAUT 1006 (October 2021)