---
title: 'Top Gas Leaks'
description: Ranks fleet metering nodes by gas-leak probability (0-100) from four consumption markers, with an adaptive per-node threshold.
section: AI Analytics
weight: 5
related:
  - ai-analytics/fleet-reports/decarbonization
  - ai-analytics/fleet-reports/suspicious-nodes
---

import Alert from '@/components/docs/Alert.astro';

A ranking of metering nodes across the entire fleet by how likely it is that a node has a gas leak. The report is for operations and the monitoring service: it suggests where to send a crew with a gas analyzer first. Each node is assigned a score from 0 to 100 — the higher it is, the more suspicious the consumption pattern.

<Alert type="warning">
This is not proof of a leak, but a probability estimate for prioritizing field visits. The final conclusion comes only from an on-site check: a pipe survey with a methane detector or an ultrasonic inspection.
</Alert>

## What it shows

The main table is a list of the most suspicious nodes with the final score and the four "marker" features it is built from. Clicking a row expands a chart of daily consumption over the period, showing how the background consumption holds up and where it crosses that node's leak threshold.

| Feature | What it means |
|---|---|
| Consumption during quiet hours | During the calmest hours of the day (when the node usually consumes almost nothing) consumption still does not drop to zero |
| Baseline consumption | The constant lower background consumption over the whole period stays noticeably above zero |
| Persistent "dripping" | For a large share of the time there is a weak but continuous flow of gas — as from a dripping joint |
| Background growth | The baseline consumption slowly rises from the start of the period to the end — the leak is intensifying |

## How it works

First, candidates are selected from the whole fleet using reference data, then for each one the hourly gas consumption history is loaded. From this history the four features above are computed for the node and combined into an overall score with weights: quiet hours — 40%, baseline consumption — 30%, persistent dripping — 20%, background growth — 10%. The scoring follows the IEA Methane Tracker methodology.

The key feature is that each node has its own personal norm. The threshold above which consumption is considered "suspicious background" is computed separately for each node as a small fraction of its own typical consumption maximum. As a result, a small school and a large trunk-line node are each compared against their own normal level, rather than against a single common figure. The "quiet hours" are also not fixed rigidly: for each node the three calmest hours of its own day are taken.

To avoid false positives, nodes with steady round-the-clock consumption (continuous production, base-load boiler houses) are deliberately damped: their "constant flow at night" signal is normal operation, not a leak.

## Run parameters

| Parameter | Default | What it means |
|---|---|---|
| Period from / to | last 30 days | Analysis window; a longer period is useful for the "background growth" feature |
| TOP-N in the final table | 100 | How many nodes to show in the final table |
| Deep-fetch candidates | 500 | How many nodes to analyze in detail after the preliminary selection |
| Parallel fetch workers | 8 | Number of parallel download threads; lower it on network errors |
| Non-zero flow threshold (m³/h) | adaptive | You can set a single consumption threshold manually; by default each node has its own |
| Utility company | from context | Filter the fleet by organization (utility / DSO) |

## How to read the result

Look first at nodes scoring 60 and above — these are candidates for a field visit. A node with a high "consumption during quiet hours" feature and a non-zero baseline consumption is the clearest signal: gas flows when there is no one to consume it. Expand the chart: a real leak looks like a steady lower line that does not drop to zero.

What to expect from false positives:

- Sites with a genuine round-the-clock load (refrigeration equipment, continuous production) may show constant nighttime consumption — this is not a leak.
- Nodes with very little data are not included in the calculation, so as not to assign them a random score.
- The CO₂ equivalent and a monetary estimate of losses are not computed here — for that there is a separate [Decarbonization](/en/platform/v3/ai-analytics/fleet-reports/decarbonization) report.