AI fostered traffic management in China
The United Nations has estimated that by 2040 almost 60% of people are living in cities across the globe and at the same time number of trucks, cars and air kilometres will double and the amount of emissions keep rising. In addition, millions of people and animals are killed in traffic – directly (accidents) and indirectly (emissions).
Solita‘s Account Director Aki Aapaoja writes about AI’s possibilities in traffic management.
Through the ages, traffic congestion has been prevented by building new roads and lanes but more than often it is a temporary cure as it triggers latent traffic aka induced traffic. It is a phenomenon where people decided to travel by car when they otherwise would not have. So basically it is about increased demand i.e., trips caused by increased supply i.e., road capacity.
In recent years, however, there has been more and more discussion that maybe data and data-driven solutions could be cost-efficient as well as long-lasting impacts on making transport greener, more adaptive, and more accessible. Being data-driven or using data doesn’t mean throwing all the existing and working traffic management procedures into the trash can but rather it is about gradual improvements alongside existing ones. Even today, 164 EB (1018) traffic data per month is generated and hence a range of potential use cases for data solutions exists.
Daunted by traffic and congestions
The challenges of transport are global but one of the worst affected countries when it comes to sustainability, congestion, and safety of transportation. Solely in China, there are more than 160 cities over 1 million people and over 254 million registered vehicles. It goes without saying that together urbanisation and motorisation as prevailing trends lead to massive adverse effects like 1500 megatons of CO2 and almost 1 million emissions-related deaths annually. In addition, the congestion is overwhelming – the average Chinese motorist loses nine days a year stuck in traffic. If nothing is done, the future of road transport does not look rosy. Doesn’t it?
Data weaves traffic management together
Together with a Chinese partner, Enjoyor co., we started a joint research pilot at Hangzhou city in Zhejiang province, China. Enjoyor is a leading Chinese traffic management company having annual revenue of over 400 million euros and it is responsible for traffic management in many Chinese cities such as Hangzhou, Fuzhou and Nanchang.
Hangzhou is the home of world-know Alibaba and over 13 million people are living in the metropolitan area, and it is a typical crowded city that is ranked top ten among China’s most congested cities. The first phase of the pilot aims at predicting traffic speeds in different road sections for the next 10-15 minutes in three minutes intervals by using historical data crunched with sophisticated machine learning algorithms. The pilot area consists of 309 road sections and 118 intersections while the used data includes several parameters like road topology, road section speeds, and signalling schemes. Data will be also enriched with traffic and weather incident data.
The overall, long-term objective of the research pilot activities is building enabling capabilities to identify traffic phenomena e.g. congestions in advance through advanced analytics and using the information for managing and optimising traffic at a regional level. Basically, it means that different trigger points indicating the possible emergence of a traffic phenomenon should or could be identified as early as possible.
The thing that makes it both interesting and challenging is the fact that issues impacting the traffic vary over time, their relations are not fixed and there are many of them. Hence, traditional management models are no longer suitable and there is a call for AI-based models which can be continuously updated or update themselves.
AI spiced digital twin with cloud and edge computing
In many cases we highlight the meaning of data, having a crystal clear business case and how important it’s to have tangible outcomes. In some of the cases we just experiment, learn and have a bit of fun. This is exactly that latter case when you have the freedom to play around – why not. We just throw trash to ideas that someone really knows the problem or there is something called business able to understand more than normal human beings. Based on the famous Agile manifesto we could just say we value the unknown more than something we think we know. We as people are sometimes biased, and look at things in a very narrowed way and especially when it comes to traffic – building a digital twin of traffic would be fun?
The challenge to build smart systems has been Conway’s law is an adage stating that organisations design systems that mirror their own communication structure. We are dedicated to trying a model where everyone in the team can time-box a bit of their study and work in parallel and together to vote for winning ideas so we would not be building a “one-size-fits” nobody solution. That was a huge success – in a few days we had already spatial data on visualisations, ongoing data replications going and tiny ML running on the edge. Now comes the interesting part – how to integrate all together for something you might see on PowerPoints as the future-proof architecture?
It’s the communication, team motivation, and keeping things small when possible so it’s easy to adapt. The team chose to use Amazon SageMaker Autopilot that can train and optimise hundreds of models based on these algorithms to find the model that is a good candidate for us. Same time few were working with AWS SageMaker to run machine learning models to find anomalies of data and anything suspicious – and immediately found time series data was fixed (typical on machine learning cases when we lack data). This incorrect way of fixing data resulted in bad models. This was not possible to be detected using any visualisation or typical data engineering tools.
After a few coffee breaks the first data API product was available on AWS serverless API development portal. Taking all this machine learning (EdgeOps) to Edge where resources are very limited can be accomplished using a rule of thumb – keep it very small and simple. Running all interference at Edge will bring few benefits like improved latency, security, and resilience. End of the week we were able to see which parts are common, automate all using AWS CDK and keep only the parts that are really needed to avoid feature creeps.
So did we build a full-blown digital twin? Not yet, and it does not matter. We found relevant feature importances from data assets and we can not wait to proceed on to the next step. Setting up even crazy goals and making experiments using new services from hyper scalers like AWS can be a first step to start something new. We now have rock solid scaleable edge, cloud, and MLOps solutions with few rock-solid models to teach us something new. Technology capabilities are outstanding and on that hype, it’s good to remember that a good team and trust in it is equally important. We encourage you to set your data and machine learning models free!
From tech to business and vice versa
For Solita, V2X is a spearhead project when it comes to holistic industrial internet of things capability development the whole funnel from mobile vehicle (car, truck, ship, machine) or stationary device (process, production) to end-user via cloud system including data pipes, real-time data processing, edge computing, and access control in some cases data farming when source data is not available yet. Technology development is not the main thing, but novelty comes from the collaboration between different people and ecosystems.
Read the original article on Solita’s web page.