By: Krista Davidson
29 Nov, 2021
Adam White, Canada CIFAR AI Chair and Amii Fellow, is collaborating with a team of researchers including Canada CIFAR AI Chairs and Amii Fellows Martha White, Csaba Svepesvari and ISL Adapt, an artificial intelligence (AI) engineering firm, to develop novel solutions to supply clean drinking water to the residents of Drayton Valley, Alberta.
White is collaborating with a team of researchers including Canada CIFAR AI Chairs and Amii Fellows Martha White, Csaba Svepesvari and ISL Adapt, an AI engineering firm, to develop novel solutions for controlling water treatment processes and supplying clean drinking water to the residents of Drayton Valley, Alberta.
While fresh drinking water is as easy as the turn of the tap for most Canadians, it has been a challenge for the town of Drayton Valley, which has seen exponential growth in its population over the years. In 2008, ISL Engineering and Land Services conducted an assessment that forecasted the need for a larger, more sustainable drinking water supply in order to serve residents and to eliminate the ‘boil water’ advisory that often afflicts the town’s water supply due to high levels of turbidity in the North Saskatchewan River. Their recommendation led to approved funding for the development of a state-of-the-art water treatment facility with sophisticated, sustainable filtration mechanisms, a treatment capacity of 18 million liters per day, and 2800m³ of additional potable water storage to meet the growing demands of the population.
How to maintain the water treatment plant to ensure it remains operational, cost- and energy-efficient is at the heart of this academic/industry collaboration.
The ISL Adapt project team built a mini water treatment plant with access to the same water supply to run experiments and to collect data that will enable a reinforcement learning (RL) agent to make real-time decisions, maintaining and cleaning the system, an energy-intensive process.
The water treatment process includes two large filtration tank systems that eliminate contaminants and have to be cleaned regularly. This cleaning process takes one filter out of commission for a period of time while the other continues to function. The research team is using reinforcement learning, a form of machine learning that trains agents through a system of rewards and incentives, to predict how often to initiate cleaning processes and streamline services that minimize disruption to the town’s drinking water. “We’re experimenting by fouling up the filters and figuring out how many sensors we need to track these operational changes in real-time. This is the kind of real-world industrial data needed to develop successful algorithms that could never be tested on a real plant because of safety and regulatory concerns,” he says.
“From a machine learning perspective, one of the fundamental problems is monitoring streams of data such as turbidity readings, pH, conductivity, temperature and other sensor data and converting them in a way that enables the RL agent to make process decisions,” says White.
The team is using policy gradient methods, a type of reinforcement learning technique that is ideal for situations that require continuous decision-making: taking actions in anticipation of long-term rewards. This allows the agent to adapt to the types of situations that might arise in a real water treatment plant, such as reacting to changes to the river water clarity, disruptive weather events or even defunct pumping systems and fouled membranes.
“There is a bit of a meme that AI has a bad reputation for taking away jobs but this project is not about replacing humans. The Drayton Valley mini plant and its RL control system can be replicated in other remote communities with water treatment challenges that lead to boil water advisories. We can’t always build large sophisticated water treatment plants nor always guarantee that there will be skilled operators to run them. If the filtration systems get fouled or contaminated water enters the treatment chain there may not be a skilled operator present to detect these problems which can result in hundreds of thousands of dollars for filter replacement costs and this can be devastating for small communities with limited budgets,” says White.
White is cross-appointed between the University of Alberta’s Department of Computing Science and DeepMind Alberta. He says that reinforcement learning is the closest form of AI to understanding natural intelligence and he is inspired by how RL might bridge the gap between the two, noting that Amii and CIFAR have been critical to advancing Canada’s leadership in AI.
“Canada has a really strong role in AI and machine learning and that’s driven by funding, support and collaboration from CIFAR over the years. I hope to see more collaborations like this in the future that aim to progress our understanding of human and machine intelligence,” he says.