Every day, the average Norwegian sends 180 litres of water into the sewage system. Whether it comes from the kitchen sink, the shower or the lavatory, the water will sooner or later end up in the sea. Fortunately, we have effective treatment plants that ensure that the water in our fjords is safe for birds, fish and bathers alike.
But wastewater treatment is a complex process. Treatment plant operators have to consider variables such as phosphate, nitrate and microparticle concentrations, and at the same time keep an eye on water volumes, temperature and alkalinity.
“Wastewater also contains a lot of infiltration from groundwater and storm water, which are both dependent on precipitation levels”, says Hilde Johansen, who is a Project Manager at VEAS, which is Norway's biggest wastewater plant. VEAS treats sewage produced by about 770,000 people living in Oslo, Bærum, Asker and Nesodden. “Infiltration results in a lot of variation in the quality of the water entering the plant, and this variation makes decontamination more challenging”, says Johansen.
This complex mix of variables means that plant operators have to make continuous adjustments to the treatment process, and Johansen admits that they could do with a little help.
“Our current control systems are insufficiently precise, and would benefit from greater optimisation. It is here that artificial intelligence and smart control models can play a role”, she says.
As part of a cross-disciplinary project called INVAPRO, process engineers at VEAS are working with research scientists at SINTEF Digital to develop an AI-based tool that will make the wastewater treatment process more efficient. The municipal water authorities in Bergen (Bergen Vann) and Trondheim are also participating in the project, which has received NOK 6 million in funding from the Research Council of Norway over a three-year period.
By utilizing all available data on the water treated in these municipal plants, machine learning methods will be employedto learn connections and calculateexactly how much air and chemicals are needed to make the essential bacteria feel at home. These bacteria convert the nitrates in the water to nitrogen gas. At VEAS, the gas is then removed in the plant before the effluent water is released into the fjord.
The smart AI-based control models will ensure that the bacteria are always provided with the best possible conditions in which to do their job, enabling them to decontaminate the wastewater more easily and efficiently than is possible using the current system.
“We’re hoping to develop control models that can precisely regulate treatment process variables according to the properties of the water entering the plant at any given time”, says Johansen.
The aim is to install new and intelligent systems that make wastewater treatment processes more efficient. For the people served by these plants, this may mean cheaper municipal rates for water and sewage treatment services.
“Such services are becoming a major national concern, plagued as they are by an ageing infrastructure of sewage networks and treatment plants”, says Anders Bryhni, who is coordinating an initiative called AI@SINTEF. The aim of this initiative is to combine research into machine learning with the application of AI in the various industrial sectors where SINTEF has its research focus.
According to Bryhni, the total infrastructure maintenance lag in the Norwegian wastewater treatment sector amounts to an estimated NOK 390 billion, which will ultimately only result in increasingly higher municipal rates for the communities it serves.
The INVAPRO project is adopting a cross-disciplinary approach to meeting this challenge. In order to make wastewater treatment more efficient, the AI experts first have to understand how sewage decontamination processes work – and this is where the expertise of the process engineers that operate the plants comes in. Based on experience exchange and data from the VEAS plant, researchers will be able to develop models and algorithms capable of regulating the treatment processes. In their turn, the plant operators will obtain more accurate indicators on which to base their work than they have access to today.
“Machine learning works in such a way that the model teaches itself on the basis of the data available to it”, says Signe Moe, who is a Research Scientist at SINTEF Digital. “In this case, such data may include the level of contaminants in a given volume of water, and the amount of chemicals used to decontaminate the water”, she says.
Can’t understand good weather
The researchers are taking their cue from historical data derived from a number of different aspects of the water treatment process. Moe, however, is eager to point out that the treatment process involves so many variables that a lot of trial and error will be needed before the AI system learns to understand what is going on.
“We previously ran a preliminary project in Trondheim municipality which is known for its poor weather, she says. “But because the model hadn’t been exposed to good weather data before, it couldn’t understand what was happening”, says Moe.
However, even though good weather generated some problems, the preliminary project yielded some promising results. The researchers succeeded in identifying the variables that have the greatest influence on the decontamination process, and which of these should be integrated into the control of chemical dosages. For example, it emerged that the treatment process is quite slow in the sense that the water in the plant is very well mixed. Water that flows out at a given time may thus have entered the plant either relatively recently, or several hours ago.
Experience from the preliminary project has formed the basis for further work. As the model gradually learns to understand and deal with the large amounts of data involved, it will be able to contribute towards boosting the efficiency of the complex treatment processes above its previous levels. This will result in considerable benefits for the wastewater plants, which will be able to exploit a potential only offered by the power of smart data. During its development, the researchers will also conduct evaluations of the model, and will be able to provide advice and recommendations on variables that ought to be included in calculations performed during the treatment process.
“We may find that improvements can be made to the model by including information obtained from outside the plant, such as meteorological data”, says Moe. “When we have established confidence in the data and the model, AI will be able to provide plant operators with an excellent basis for their decision-making. Alternatively, we could cut right to the chase and let the model take autonomous control. We don’t know exactly how this will end, but our aim is to find out how machine learning can improve wastewater treatment operations”, she says.
Even though the project is still in the starting phase, the partners have a lot of faith that INVAPRO will achieve its aims.
“We’re hoping that it will provide us with better and more stable wastewater treatment processes, as well as reductions in chemical use and energy consumption”, says Hilde Johansen at VEAS. “If we succeed, both operational stability and safety will increase significantly, and perhaps we may be able to postpone infrastructure renovation”, she says, going on to add that VEAS also recognises a potential for in-house skills development when it comes to processes.
The participation in the project of VEAS, Bergen Vann and Trondheim municipality means that the technology can be tested on a number of different types of treatment plants. If the model succeeds in meeting the challenges faced by these plants, the partners will also be able to apply the technology in most others.
Researcher Signe Moe believes that the INVAPRO project partners are pioneers in the field of trying out machine learning in connection with wastewater treatment.
“It’s very likely that other groups are working on this too, but so far no products that integrate machine learning have appeared on the market”, she says. “We are also seeing synergies being generated with other research activities going on at SINTEF. For those of us working on enabling technologies and methods, the process being described by the data is not so important. Experience from the INVAPRO project can be applied in a broad range of industries and other fields”, says Moe.
His colleague Anders Bryhni backs her up.
“Machine learning offers great potential in the field of automated control in process plants, especially when it comes to safety-critical processes such as the regulation of smelting furnaces and oil wells”, he says. “We can learn a great deal from trialling new things in connection with wastewater treatment. The knowledge we obtain can be applied in other critical fields where the consequences can be catastrophic if things go wrong”, says Bryhni.