How this utility is fixing water main failures before they happen
by Water Source
Machine learning is helping mitigate the impact of the 44,000 water main failures that happen across Australia each year.
Attempting to predict when a water main will fail is a costly exercise that generally requires visual inspection – not an easy feat when the asset is underground. But waiting to act until a main fails also has flow-on consequences.
“Each service interruption is an inconvenience for our customers,” said Western Water General Manager for Strategy Livia Bonazzi.
“It’s also an extra cost for the emergency repairs, not to mention the water losses.”
https://omny.fm/shows/australianwater/dilusha-weeraddana-livia-bonazzi-on-machine-learni
Western Water, which services 160,000 people in the northwest of Melbourne, experiences about 400 water main failures a year. With the region’s population set to double over the next 15 years, Bonazzi said it is important to look at how to do things differently.
“We’re grappling with the challenge of how to service a growing customer base using the most efficient and cost-effective solutions,” she said.
“It made me think about what it would be good to have under the Christmas tree that would help us with this.”
Ideally, Bonazzi said the utility wanted a model that could detect the moment just before a pipe failed, allowing it to preemptively repair or replace the asset.
Thanks to a chance meeting with CSIRO’s Data61 team at Ozwater’18, this is what Western Water now has.
The Data61 team combined Western Water’s historical water main failure data with its data science and engineering approach to provide the utility with a 20-year forecast of when its mains will fail.
Data61 Researcher Dilusha Weeraddana said the forecast doesn’t just provide the likelihood of failure for a suburb, but for each pipe in detail.
“This wasn’t an easy task because pipe failures are not very common,” Weeraddana said.
“Plus the failures don’t only depend on age. There are new pipes that may fail earlier than older pipes as some older pipes are more robust.”
To develop the system, Weeraddana said it was a case of “feeding” a machine learning algorithm with historical data.
“We clean the data and then apply machine learning algorithms to train our model,” she said.
“It trains itself using the historical data and then predicts the future.”
Bonazzi said the tool has given Western Water robust information it can use to make risk-based investment decisions.
“The projections of the model show that by 2030 our bursts will increase by 22% and our fittings failures will increase by 26%,” she said.
“Having this forecast down to the pipeline asset level is assisting us in prioritising our water mains renewal program.”
Want to find out more? Read the technical paper here.
