Calgary water main breaks
Where Calgary's water mains break, which pipes fail, when in the year, and a model that ranks communities by risk. The City publishes breaks 1 to 3 months late.
Where water mains break, by community
Water main breaks per kilometre of active pipe since 2017, for each Calgary community.
The scale stops at the 95th percentile so one outlier does not wash out the rest. Grey: no breaks since 2017 or no pipe data. Hover a community for its breaks, pipe length, median pipe age, main material and whether the model put it in its top 20% for risk.
Water main breaks per year
Recorded water main breaks in Calgary each year since 1960; the latest year is partial.
Breaks recorded each year. 2026 is a partial year (City data through Aug 23, 2026; the City adds breaks 1 to 3 months late). The records start in 1956; the years before 1960 are sparse and left out. The trend mixes real pipe stress with seven decades of changing record-keeping.
Why water mains break: material, season and age
Breaks by pipe material and by month, and the pipe in service by decade installed.
Cast iron and other older materials carry most of the breaks; freeze-thaw months bring the most; 51% of the pipe in service went in from 1990 on.
By material of the pipe that broke
Each break counted once, on the pipe it sits on: 15,237 of 37,558 breaks, the ones on pipes still in service. Another 22,255 were on pipes since replaced, whose material the records do not give, so older materials are under-counted here.
Dark bars: cast iron and asbestos cement, the materials the risk score counts as high risk.
By month, since 2010
Breaks in each calendar month since 2010. Calgary's freeze-thaw winters bring the late-fall and spring peaks.
Dark bars: October to March.
Pipe in service by decade installed
Kilometres of water main in service today, by the decade it went into the ground.
Communities ranked by water main breaks
The highest break rates per kilometre since 2017, next to the model's top 10 by predicted risk.
Most-failing communities (observed)
By recent breaks per km of pipe (2017–present). Small networks (< 5 km) excluded.
Highest-risk communities (model-ranked)
Composite score from pipe age, material, history, training breaks, repeat-break locations, and freeze-thaw share.
The two lists overlap a lot but not fully: the model weighs pipe age and material with break history, the observed list counts breaks alone.
The water main risk model
A hand-weighted community risk score, set on 2017 to 2021 breaks and tested on the 708 breaks from Jan 2022 to Feb 2026.
The score weighs pipe age, material risk, historical and recent break rates, repeat-break locations and freeze-thaw share. The test breaks were all the City had published when the model ran on May 9, 2026; the numbers below are from that run.
Capture rate
The share of the 708 breaks recorded Jan 2022 to Feb 2026 that fell in the top 10%, 20%, 30% and 50% of communities by predicted risk (model run May 9, 2026). Random ranking would put 20% in the top 20%.
Spearman ρ between predicted-risk score and observed break density: 0.768 across 219 communities.
What the score leans on
The weight of each feature in the composite score, set by hand. They are not random forest importances; that model uses an overlapping feature set with its own.
Plain features. The 2015 UBC Bayesian model used a similar set and did about as well: well-trodden ground.
The top 20% of communities by score hold 39.7% of the Jan 2022 to Feb 2026 breaks; the top 50% hold 88.3%. Researchers at UBC published a Bayesian failure model on the same data in 2015 (Kabir, Tesfamariam, Sadiq et al.).
The satellite test: does ground movement predict breaks?
Odds of a later break per pipe, for 54,531 pipes older than the satellite window.
Ground movement measured Jul 2016 to Jul 2020 under pipes that were already in the ground, then the 1,042 breaks recorded afterwards, holding pipe age, material, length and break history constant. Ground movement added nothing; a pipe's own break history and old cast iron did.
The story behind the model
Weeks of trying to make NASA satellite ground movement predict water main breaks.
I spent weeks trying to make NASA's satellite ground-motion data predict these breaks, including one false alarm my own checks caught before publication. Read the full story →
How I did this: sources, method and limits
How breaks were placed on communities and pipes, how the risk score is weighted and tested, and its limits.
Sources
- City of Calgary Water Main Breaks (dataset dpcu-jr23): 37,558 records, 1956 to Aug 23, 2026, fetched Oct 31, 2026.
- City of Calgary Water Main Network (dataset w6h9-w33i): pipes with material, diameter and install year.
- Risk-model code on GitHub.
Method
- Breaks are placed in communities by point-in-polygon against the City's community boundaries.
- Breaks are counted per community over a 2017 to 2021 window. The risk score has six weighted features: pipe age, material risk, pre-2017 break density, 2017 to 2021 break density, repeat-break locations and freeze-thaw share.
- Tested against the Jan 2022 to Feb 2026 breaks, held out from the weighting (model run May 9, 2026, not re-run since). Spearman ρ = 0.768 between predicted score and the observed test break rate.
Limits
- Break-type codes (158 distinct strings) are City internal classifications without published documentation.
- Break records carry no pipe material. The material panel gives each break the material of the pipe it sits on (within 10 m; the City places each break on its pipe). That works only for the 15,237 breaks on pipes still in service. The other 22,255 were on pipes since replaced, and the pipe there now is the replacement, so they are left out rather than guessed.
- The risk score is hand-weighted, not learned. A separate random forest on the same features reaches AUC 0.849 ± 0.033; the deep dive has that and what happens when satellite data is added.
- This is exploratory work by a non-specialist with a lot of AI help, not peer-reviewed.
Civil engineer or pipe-failure researcher, and something looks wrong? Reach me on LinkedIn, through the GitHub repo or by email. I want the pushback.