Thames Water discharge findings
Isaac Kirk, in relation to #24.
The Thames Water data is the most useful source I've worked with on this project. It gives me start and stop times for every monitored overflow event from April 2022 to April 2026, on a network that grew to nearly 500 monitors. The full notebook is at notebooks/Thames-Water/thames-water-analysis.ipynb. This document pulls out the parts that I think matter for the website and for the next step, which is applying the same idea to the other water companies.
What I can ask of this data
Three questions, one chart each.
- Are spills going up or down year on year?
- When across the year do they happen?
- Has the typical spill got shorter, or just rarer?
Before any of that, I have to check the network is stable enough to compare years against each other.
Network coverage
The network roughly doubled between 2022 and 2024, from 328 monitors with at least one paired discharge to 494. It then held effectively flat at 458 in 2025. 2022 and 2026 are partial years. The window where year on year comparisons are sound is 2023 to 2025.
This matters because if I compare 2022 to 2024 directly, I'm mostly looking at more monitors finding more spills, not Thames Water behaving differently. Once I cut to 2023 onwards the network is stable enough that count changes mean real changes in behaviour.
Annual totals
2024 was the heaviest year on record at 29,861 paired discharges. 2025 fell to 13,464. That's more than a halving on a network that was essentially the same size (458 monitors in 2025 vs 494 in 2024).
The drop is too large to be explained by the small change in coverage. Either Thames Water's overflows really did spill less often in 2025, or 2025 was a drier year, or some mix of both. This chart on its own can't tell me which. The regression in #22 and the rainfall metric in #30 are where that question gets answered.
Monthly intensity
One line per year, all 76,848 paired events. The shape of each line is the seasonal pattern within the year. The vertical spacing between lines is the change between years.
Every year has the same shape. Spills dip through summer, climb through autumn, peak in winter. That seasonal pattern is what I'd expect if rainfall is the main driver, which the rainfall metric notebook then confirms.
The interesting bit is the gap between 2024 and 2025. 2024 sat well above 2023 from August onwards and produced an exceptional autumn into winter spike. 2025 stayed below 2024 across most months and never produced that autumn spike. So the headline drop in the annual chart is really an autumn story.
Spill duration
Boxplot of duration per spill, restricted to 2023 to 2025. The median, the middle 50 percent and the 90th percentile all hold close to constant across the three years.
So whatever changed between 2024 and 2025, it changed how often discharges happened, not how long each one lasted. The mean sits well above the median in every year, which is the right skew signature. Most spills are short. A small number run for days.
What this means for the rest of the project
Three things matter for the next steps.
First, the Thames data shows the project is worth doing. There's a real story in the count drop between 2024 and 2025 and it's too clean to ignore.
Second, the same analysis pattern works for every other water company we have EDM data for. Count active monitors per year, count discharges per year, plot the monthly shape, look at duration distributions. The Thames notebook is the template. The other companies don't have the same minute level start and stop data, but the EDM annual returns give us spill counts and total hours per monitor per year, which is enough for the first three of those four questions.
Third, none of this controls for rainfall. The seasonality of the monthly chart is too clean to ignore weather as a driver. The headline number from this notebook, "Thames Water halved their discharges between 2024 and 2025", isn't a finding worth publishing without the rainfall regression behind it. The regression in #22 and the rainfall metric work in #30 together close that gap.
Website candidates
For the three visualisation budget on the GitHub Pages site, the two strongest from this notebook are the annual totals bar chart and the monthly intensity line chart. They answer two different questions cleanly and they tell the story the rest of the project develops. The duration boxplot is the right finding (typical spill length didn't change) but a one line sentence in the text carries it without needing the chart.