What it estimates
For every water main, gravity sewer and rising main in your network, the model estimates the probability of at least one failure in the next year and over a planning horizon you choose (for example 20 years). Assets are then grouped into four risk classes, low to very high, using thresholds you set.
The model
Failures are modelled as a Poisson process. The failure intensity per kilometre of main follows a Weibull power law in the asset's age, so failures become more frequent as pipes get older, at a rate that depends on the pipe's cohort.
expected failures per km between ages a0 and a1 = (a1 / eta)^beta - (a0 / eta)^beta
P(at least one failure) = 1 - exp( -length_km x expected failures per km )
Parameters (beta, eta) are held per cohort: material, diameter band and install era. Longer mains are more likely to fail than short ones. Repairs do not reset a main's age.
Calibration to your network
Out of the box the model has illustrative default parameters. They are not fitted to any real network. The value comes from calibration, where the parameters for each cohort are fitted to your failure history by maximum likelihood.
- Each cohort needs a minimum number of recorded failures (30 by default). Cohorts with too few failures fall back to a broader group (material and diameter, then material alone), and finally to the defaults, so a thin history never produces a spurious fit.
- The fitted parameters are reviewed and approved before use. A calibration belongs to your organisation, and you can see which one each run used.
Backtesting
Before a calibration is approved, the most recent years of failure history are held out. The model is fitted on the earlier years, asked to predict the held-out years, and compared with what actually happened. We report:
- Brier score and skill against a constant-rate baseline. Skill above zero means the model beats simply assuming every pipe is equally likely to fail.
- Top-decile capture: the share of assets that failed in the held-out years that were ranked in the model's top 10%. A random ranking captures about 10%.
- Calibration by decile: predicted against observed failure rates across the ranking.
- Expected against observed failures over the held-out period.
Explainable results
Every asset shows the factors behind its result (its age, its cohort, and its length), with how much each raises or lowers the risk, so a renewal decision can be explained to a board or a regulator.
Reproducible and versioned
- Every run records the model version, the calibration used, the resolved parameters and a fingerprint of the input data. The same inputs give the same results.
- Model versions move through a defined status flow (development, validated, approved, deprecated), and released versions are immutable. Your results don't change between reporting periods unless you choose to upgrade.
- Each model version has a published model card describing its purpose, method, assumptions, validation and known limitations.
Assumptions
- Failure intensity depends on age, cohort and length only. Soil, pressure, operating conditions and other covariates are not modelled in this version.
- The recorded failure history is reasonably complete, and its dates are accurate, within the calibration window.
- Install year is the start of the asset's life. Where an install year is missing and inferred from the cohort median, that asset's result is marked as indicative.
Known limitations
- This version does not estimate remaining useful life or the consequence of failure. Those models are in development.
- Cohorts with too few failures fall back to default parameters that are not fitted to your network.
- Networks with very short failure histories give wide parameter uncertainty. Confidence intervals are not reported yet.
- Condition data (for example CCTV grades) is accepted but not yet used by this model.
- Results are decision support. Renewal decisions remain the responsibility of the asset owner.
Validation to date
Version 0.1.0 has been tested on synthetic networks generated from a known failure process: it recovers the true parameters within tolerance, and its backtest beats the constant-rate baseline. Validation against real utility networks takes place as part of each pilot, using your held-out failure history, and the results are shared with you.