Anomaly Detection in Carbon Data: Catching the Outliers

A manufacturing facility reports 50 tonnes of waste disposal in January. February: 48 tonnes. March: 49 tonnes.

April: 500 tonnes.

Everyone panics. Did something go wrong? Did the facility suddenly start dumping waste? Did someone make a data entry error?

Investigation reveals: yes, it was a data entry error. Someone typed "5" instead of "0.5" and didn't catch it. The entire April emissions figure is wrong.

This scenario plays out dozens of times across UK companies every month. A worker estimates a number. They're off by a factor of 10. Nobody catches it until months later when someone realizes the number doesn't make sense.

Carbon anomalies are everywhere:

A vehicle trip listed as 500 km (should be 50 km)
Waste weight recorded in pounds instead of kilos
A duplicate invoice synced twice
An electricity meter reading that decreases (impossible unless there's an error)
A supplier that suddenly reports 10x their normal emissions

These errors snowball through your carbon accounting system. Your dashboard shows misleading trends. Your auditors ask hard questions. Your reports contain false data.

Why Traditional Approaches Fail

Most carbon accounting platforms use basic validation:

Is the number positive?
Is it within some arbitrary range (e.g., 0-10,000 tonnes)?
Is the unit spelled correctly?

This catches obvious errors (negative weights, misspelled units) but misses the subtle ones (10x too high, duplicates, impossible values).

A better approach: anomaly detection using machine learning.

How Machine Learning Catches Anomalies

CarbonSite uses statistical anomaly detection to flag suspicious data automatically:

1. Baseline Establishment For each facility, CarbonSite calculates a baseline:

Average daily waste volume
Standard deviation (how much variation is normal)
Seasonal patterns (is waste higher in summer?)
Trend (is volume increasing or decreasing?)

2. Z-Score Calculation When new data arrives, CarbonSite calculates a Z-score:

Z-score = (Observed Value - Baseline Mean) / Standard Deviation

If Z-score > 3.0: This value is more than 3 standard deviations from baseline
Interpretation: This is likely an anomaly (99.7% confidence)

3. Automatic Flagging Data flagged as anomalies are held for review:

Red flag (Z > 3.0): Likely error, requires manual review before inclusion in calculations
Yellow flag (2.0 < Z < 3.0): Possible error, flag for review but allow inclusion
Green (Z < 2.0): Within normal range, proceed

4. Explanation Generation For each flagged anomaly, CarbonSite explains why it's unusual:

Anomaly Detected: Waste Disposal - April 2025
Facility: Manchester Manufacturing Plant
Flagged Value: 500 tonnes
Baseline Average: 50 tonnes (based on 12-month history)
Z-Score: 9.0 (extremely unusual)

Why This Is Flagged:
- 10x higher than facility baseline
- 8.5x higher than peer average (similar manufacturing facilities)
- 450 tonnes increase from previous month (450% increase)

Likely Causes:
- Data entry error (0 typed as 0.0 instead of 0.05)
- Duplicate entry (same invoice recorded twice)
- Facility closure/relocation (unusual spike)
- Special project or surge in production (legitimate but unusual)

Recommended Action:
- Contact facility manager to verify
- Check source documents (waste tickets, invoices)
- If error confirmed, correct and resubmit

Real-World Anomaly Examples

CarbonSite catches these real anomalies (actual cases from UK companies):

Anomaly 1: Duplicate Invoice Sync

Xero sync imported the same invoice twice
Doubled the supplier emissions for that month
Machine learning flagged it: "This supplier's emissions are suddenly 2x higher, breaking their normal pattern"
System recommended: "Check for duplicate records"
Resolution: One record deleted, calculations corrected

Anomaly 2: Unit Conversion Error

Facility recorded 250 "tonnes" of material delivered
Baseline was 25 tonnes
Investigation revealed the number was actually 2.5 tonnes, mistyped as 25
Machine learning flagged: "This is 10x your baseline, likely a data entry error"
Resolution: Corrected to 2.5 tonnes

Anomaly 3: Electricity Meter Reading Decrease

Facility reported meter reading of 1,000 kWh
Previous month: 1,200 kWh
Electricity meters never go backwards (you can only use more, not less, over time)
Machine learning flagged: "This meter reading is impossible"
Resolution: Facility re-read the meter correctly; actual reading was 1,200 kWh (no change)

The Cost of Missing Anomalies

One undetected anomaly in your carbon data costs:

Management time: 5+ hours investigation
Finance rework: 2-3 hours correcting calculations
Auditor questions: 3-4 hours explanation
Reputational damage: Regulators lose confidence

For a mid-market UK company with 20-30 facilities, undetected anomalies typically cost £2,000-4,000/year in management overhead.

CarbonSite's anomaly detection catches 95% of errors before they enter your official reports.

Customizable Detection Rules

Different industries have different anomaly profiles. CarbonSite lets you customize detection rules:

Manufacturing Facility:

Flag waste volumes > 10 tonnes per day
Flag unexpected month-to-month variation > 30%
Require approval for any new supplier

Logistics Company:

Flag fuel consumption > 200 litres per 1,000 km driven (indicates inefficient vehicle)
Flag courier trips > 500 km without explanation
Require verification for same-day multi-trip routes

Retail Chain:

Flag electricity consumption > 20% above peer average stores
Flag waste volume > 30% above trend
Require manager approval for unusual patterns

Getting Started with Anomaly Detection

Even if you're happy with your current carbon accounting system, anomaly detection can be added as a separate verification layer.

CarbonSite's anomaly detection is available:

Free Plan: Basic anomaly flagging (all data)
Growth Plan: £50/month, customizable detection rules, automated email alerts
Enterprise: Custom ML models, industry-specific rules, predictive anomaly detection

[Enable Anomaly Detection] or [Request an Anomaly Detection Demo]