New machine learning tool predicts sewage overflows
A new AI tool predicts when sewage overflows will occur during extreme wet weather events, helping utilities potentially avoid them completely.
The tool, developed by Unitywater and GHD Digital’s AquaLAB, uses machine learning to accurately predict wet weather impacts faster than traditional hydraulic models.
It does this by understanding the link between past rainfall imagery from the Bureau of Meteorology and whether an overflow occurred. It then applies this knowledge to new, live radar images to generate a prediction up to six hours in advance.
Unitywater Asset Performance Engineer Gagneet Serai said a review of the disruption caused by Cyclone Debbie in 2017 prompted the utility to look for a different way to forecast sewer overflow events.
“Cyclone Debbie had a big impact; there were a lot of overflows,” Serai said.
“Our performance was pretty good, but there was room for improvement.”
To improve, the utility needed to better understand risk and make decisions, which is how Serai and GHD Digital’s Data Scientist Jeff Fisher landed on machine learning. The pair will discuss the tool and the collaborative way it was created at Ozwater’19 on Wednesday 8 May.
Intuition meets machine
When designing the tool, Serai and Fisher wanted something complex enough to capture the distribution of rainfall that could also replicate human intuition.
“Operations staff who have been with a utility for decades have a gut instinct,” Fisher said.
“They know if a big storm comes from the southwest which pump stations will flood … We wanted to build a model that was similar to the intuition of these skilled operators.”
They also wanted to make it easy to use, which meant presenting the information in a visual way rather than as a sheet of numbers.
“Looking at a page of numbers isn’t as easily accessible as looking at a map with pump stations flashing red and saying this is critical,” Fisher said.
“That was the other aspect of this project: now that we’ve got a model that can produce these probabilities, how can we get that information to someone as quickly and as easily as possible?”
For this to work, Serai said Unitywater and GHD needed to collect vast amounts of data and radar images. The high quality of this data then enabled the development of a reliable prediction tool.
“We need to appreciate the importance of having accurate data, because until you start using it you don’t understand how important it is,” he said.
“Before you jump into machine learning, you need to know if you’re capturing the data that you need and are you capturing the quality that you need.”
Find out more about Unitywater and GHD’s overflow prediction tool at Ozwater’19 from 7-9 May. To view the full program, click here.