Every water company publishes interruptions. None publishes a comparable view.

Ofwat measures supply interruptions as minutes lost per customer, for events of three hours or more. The companies publish the underlying event files. They use different field names, different category vocabularies and different portals — so nobody can put them side by side. Normalising them is the work, and this is what the first one looks like.

Yorkshire Water, 2018–19

Event rows	1,899
Distinct interruptions	1,500
Properties affected	83,528
Median duration	4.9 h
90th percentile	9.8 h
Distinct DMAs	646

Derived aggregates from the company's own published file. Every event is ≥ 3 h because that file is the performance-commitment dataset, which counts only those.

What normalisation actually means

Four category names for two things

The company's words

UNPLANNED
PLANNED & WARNED
PLANNED TO UNPLANNED
UNPLANNED BUT CAUSED BY 3RD PARTY

One file, four vocabularies, no definitions inside it.

The canonical form

planned — 90 rows
unplanned — 1,809 rows
unknown — held for review

The original wording is kept beside every normalised value, so nothing is lost and the mapping can be audited.

A second company will invent a fifth vocabulary. That is why the mapping is data, not code, and why a changed header stops the import instead of guessing.

Why this is not a live map

The live feed is theirs. The comparison is nobody's.

A real-time interruption map needs a water company's own systems. What is publicly available is the published event history, and its value is that it can be compared: which DMAs repeat, whether events are planned or unplanned, how durations differ between neighbours, whether a company's own wording hides a category it would rather not headline.

The pipeline behind this page keeps provenance on every row and refuses to import when a portal changes its columns — because a data product that silently mis-maps a field is worse than no product.

State of the sources

One works, four are walled

Reachable

Yorkshire Water via Data Mill North — a direct CSV, 1,899 rows.

Attempted and blocked

Ofwat — the performance-report publication page returns "You can't access this page right now" even to a headless browser. DiscoverWater — the interruption URL returns 404; the site is a client-rendered application with no exposed data endpoint. data.gov.uk — the national catalogue holds one interruption dataset (Yorkshire, via Data Mill North) and nothing from the other companies.

That gap — files that exist but cannot be found or fetched together — is the reason a normalised view is worth building, and the reason it is slow.

Six years, one measure

Minutes lost per property, 2014–2020

The same company publishes its performance commitment values by year. Supply interruptions (commitment WB2) is the Ofwat measure: minutes lost per property, for interruptions of three hours or longer, planned or not.

Year	Published value	Minutes
2014–15	09:29	9.5
2015–16	12:53	12.9
2016–17	09:47	9.8
2017–18	06:58	7.0
2018–19	10:28	10.5
2019–20	07:34	7.6

The 2018–19 event file and the 2018–19 commitment value describe the same year from two directions: 1,500 interruptions in the events, 10.5 minutes lost per property in the commitment. That two-grain check is what makes an observatory credible — an event without a metric can be an anecdote, and a metric without events cannot be inspected.

Method

What was done to the data, exactly

Geometry

OSGB eastings and northings converted to WGS84 through the published Ordnance Survey route and Helmert shift, in plain Python. Checked against the OS worked example to about a metre.

Validation

Every row checked for required fields, a parseable start time, a plausible duration, non-negative properties, and coordinates inside Great Britain and Ireland.

Provenance

Source URL, retrieval timestamp and source row number travel with every record. Rejected rows are kept with their reason rather than dropped.

Source: Yorkshire Water performance commitment 6 file, published via Data Mill North. The figures here are derived aggregates, not a reproduction of the dataset, and the source is linked rather than redistributed. This page is a demonstration of method, not a regulatory finding about the company named.