The intake of fresh water is under pressure in more and more places in the Netherlands. Due to persistent drought and rising sea levels, salt water is penetrating further into the country, with far-reaching consequences for nature, agriculture and drinking water supplies. DigiShape partner HKV Lijn in Water has traditionally been involved in these issues and advises governments and partners on appropriate measures and future scenarios. Although 3D models offer a lot of insight in this regard, they are often too slow for situations where quick action is required. That is why HKV is exploring the use of Machine Learning as a supplement to the traditional approach. Vincent Vuik and Thomas Stolp of HKV explain it in this interview.
“It seems to be becoming the new normal that increasing drought is causing problems in the intake of fresh water,” says senior advisor coastal and hydraulic engineering Vincent Vuik. “Since 2018, we have seen a pattern: a normal summer, followed by an extremely dry summer, then a normal year and then drought again. This variety continues, salinization is no longer an exception. This trend is leading to more and more questions from governments and water managers. “What should we do with this in the short, medium and long term? Will the entire port soon be closed off, so that no more salt comes in? These kinds of questions are alive in society, which is why we carry out extensive model studies and system analyses in all kinds of projects. In this way, we help parties to make well-founded decisions.”
3D models for salt intrusion and associated measures
An example of such a project is the modelling of salt intrusion in the North Sea Canal, where HKV contributes to better water management with system analyses and scenario explorations. 3D water models are also being used in the Rijnmond and the Kiel Canal to gain insight into the distribution of salt under varying conditions. An important question is often what measures are needed to prevent further salinization.
“One solution that we are now exploring with a contractor is a temporary threshold of sand in the river,” says Vincent. “It may stop the salt water, but it also hinders shipping. That is why we calculate in advance exactly what such a threshold would yield. Does it really help, or does the salt still slip over? And where do you put it: close to the sea or further inland? How big should it be, how wide, in which branch of the river? We answer such questions with model calculations, so that the contractor can take well-founded follow-up steps.”
The added value of Machine Learning in addition to 3D models
“3D models are extremely reliable for this kind of calculation, but with Machine Learning you can go one step further,” says AI specialist Thomas Stolp, who works on the use of artificial intelligence in water projects at HKV. “Numerical models are often purely physical and do not take into account, for example, shipping or discharge management, while those factors do have an influence. If you still want to include those kinds of processes, you have to expand the entire physics in the model, which takes a lot of time and computing power. With Machine Learning, you can integrate these influences relatively easily, by adding them as features. This enriches the model, without making it more complex or heavier.”
What is Machine Learning?
“Machine learning is a method in which you don’t explicitly prescribe everything to the model, but let it learn from examples,” Thomas explains. “You feed the algorithm with data, for example from your 3D models of salt intrusion, and based on that, it learns to recognize connections and make predictions itself.” This is much faster than with traditional models. In the event of acute drought, you sometimes want to know within an hour what the effect of a measure is, such as temporarily opening a sluice. “That is hardly possible with a 3D model, but with a trained ML model you have an estimate in seconds. Conversely, you can also calculate thousands of climate scenarios in a short period of time, something that would take days or even weeks with traditional models. In this way, we not only make our analyses faster, but also more widely applicable.”
Two directions: forecasting and broadening
According to Thomas and Vincent, there are currently roughly two ways in which Machine Learning can be of great value. “On the one hand, you can use it for short-term forecasting,” says Vincent. “For example, to quickly determine whether a collection point will still be usable in a few days.” On the other hand, Machine Learning offers opportunities to explore scenarios that you simply cannot calculate all with traditional 3D models, according to Thomas. “We now often calculate with large steps, one or three meters of sea level rise, or fixed river discharges, because full simulations are too slow for that. But it is precisely between those scenarios that valuable information is found. What if a storm suddenly comes up? Or a ship passes?” With Machine Learning, you can still include these kinds of variations, without having to simulate each scenario separately. This creates a smart layer around the physical model: a hybrid approach. You maintain the reliability of the 3D calculations, but add the speed and flexibility of AI.”
Confidence in the model thanks to Explainable AI
In AI applications, it is not only important that a model makes a prediction, but also why. That is why HKV is focusing on Explainable AI: techniques that provide insight into how the model arrives at a certain outcome. “If you advise a measure that has consequences for the drinking water supply or shipping, as an advisor you want to be able to explain what you are basing that prediction on,” says Thomas. “Transparency is essential. That’s why we’re working on models that are not only fast, but also explainable.”
The use of AI in water management is still in the pioneering phase, but anyone who still wants to keep their feet dry in fifty or a hundred years’ time should start investing now, says Vincent. “You can’t wait until the need arises. Now is the time to build the models of the future.”
Taking the next step together
HKV combines decades of experience with physical models with up-to-date knowledge of AI and Machine Learning. This combination is essential to develop reliable and applicable solutions. “We know how the system works, how you can capture that system in data and how we can innovate our calculation models,” Thomas and Vincent conclude. “We are happy to share this combination with Rijkswaterstaat, provinces, water boards, port companies, drinking water companies, engineering firms and contractors. Let’s explore together how AI can strengthen the water management of tomorrow!”
More information
Also read the informative background article Fresh water under pressure: how AI helps with smarter water management on the HKV website. For more information, please contact Vincent Vuik or Thomas Stolp.
