The wheels are being set in motion at the HHLA Container Terminal Burchardkai (CTB). On the 1.4-square-kilometre site, container gantry cranes are moving containers between ship and land, straddle carriers are transporting them to the container yard or rail terminals, and trucks are moving in between with their container loads.
It can look like a commotion, but all processes are perfectly coordinated. Nevertheless, the port logistics staff would be happy to avoid some of this motion: “We have to move some of the containers more often than strictly necessary, because we do not have the full information about when and how they’ll be leaving the terminal,” explains Christian Greinert. As an expert in terminal development, he oversees the CTB’s storage management.
“We call these containers restackers. In some circumstances, they have to be moved around the yard often or transported over long distances in the terminal,” adds his colleague Nico Marks. “If we had the full data this wouldn’t be necessary.”
This extra work costs time and storage capacity – both of which are scarce. To solve this problem, Greinert and Marks have developed a milestone in port logistics, together with the support of external experts. The gaps in information in the freight documents are filled with the help of artificial intelligence (AI) at Burchardkai, optimising container stacking at the terminal – with a surprising level of accuracy.
For 50 percent of the containers that are delivered by ship it is unclear on arrival when they will leave the terminal again. For every tenth container it is also unknown whether it will be moved to its destination by waterway, rail or truck. And it is precisely this information that is decisive for the optimum terminal operation because it determines where the container will be stored until it is transported on.
“The target for port logistics is to get containers to the optimum storage space as quickly as possible and to get them out again just as quickly,” adds Greinert. The aim is to achieve the shortest possible journeys for the fleet of 133 straddle carriers and as few restacks as possible in the largely automated storage blocks, where up to five containers are stacked on top of one another.
In order to achieve this aim, Greinert and Marks have introduced a machine-learning module in the integrated terminal management system (ITS). Without knowing how a container will be transported on or its dwell time, ITS was only able to assign the container a space at random until now. Whether it was the ideal space would only become clear once the container was removed.
“In some cases, the container had been transported twice over far too long distances,” clarifies Marks. Thanks to the ML module, the management system now has an overview – the AI can forecast how and when the container will leave the port using more than 30 different datasets/information about the container.
In addition to the origin, container shipping company and the course of the transport so far, information about the weight and final destination are also among the data analysed. The system compares this information with master information gathered from hundreds of thousands of container movements on the terminal in the past and looks for similarities.
“Because the system can gather and analyse a lot of data in a very short amount of time, it recognises patterns and correlations that we as humans would not even be able to guess at,” explains Nico Marks. A fictitious example is given of two containers being transported to Aarhus. If the destination and weight for both containers are known and the same, the AI is able to derive a pattern based on the data available to it that forecasts with a high level of certainty one container being transported on by truck and the other by feeder. “In the past, we simply assumed based on the destination that both would be loaded onto a ship,” says Greinert.
A pilot project based on AI was launched at the HHLA Container Terminal Altenwerder at the end of 2019 to determine dwell times. CTB was involved at the beginning of this project, but due to the specifics of the Burchardkai it decided to go its own way. The CTB model was based on a concept by HPS Hamburg Port Consulting – a HHLA subsidiary – and the software company INFORM GmbH from Aachen.
Work on the module took a good year. During this time it was trained with 1,000,000 datasets from container handling over the last two years for deployment at Burchardkai. The system was able to grow its knowledge with three quarters of this information – the remaining 250,000 datasets were used for validation to check whether this knowledge was being used correctly
The machine-learning process was a success: the system is able to determine the correct dispatch area (quayside or hinterland) for 77.5 percent of the containers lacking information. As a result, containers were stored in a better location in the terminal. The rate of restacks – which are unavoidable even with full information – decreased by 8 percentage points. The share of containers stored optimally according to the time when they will be dispatched has increased from 57 percent to 70 percent.
The efficacy can also be seen in the detail: the distance covered by straddle carriers loading a feeder ship were shortened by just over 25 percent. “With these savings we have recovered the project costs in less than half a year already,” says a pleased Greinert.
The collaboration with HPC and INFORM GmbH as well as colleagues at the Container Terminal Altenwerder “was a significant factor in our success,” explains Greinert, “because our partners brought with them a profound understanding of the operating processes.” However, the product’s solid performance is not the only outstanding result that this project has delivered according to Greinert and Marks: “We were able to integrate the ML module directly into the terminal management system without major changes.”
This success will radiate out beyond this terminal: “We have shown in practice what can be achieved with artificial intelligence in logistics and port operations.” Christian Greinert and Nico Marks will continue developing and refining their system. The next step will involve updating the database that the module uses to make its decisions.
Greinert and Marks are also considering letting the ML module make decisions in other areas of the terminal management system where the database is convoluted or not accurate enough. Greinert and Marks mentioned order management for the automated block storage cranes, the arrival forecast for external trucks, and the possibility of a yard forecast for strategic yard planning as examples.