Anomaly Detection in Carbon Data: Catching the Outliers

The Data Entry Error That Cost a Company a Week

A waste contractor submitted delivery tickets for Q3. One ticket showed:

367 tonnes of waste captured in a single day

The facility's historical average: 40 tonnes per day.

Nobody caught it until the auditor asked: "Why did this facility's emissions jump 8x in week 2 of September?"

Investigation revealed: Field worker entered "367" (meant to be 36.7 tonnes, decimal point missed).

Impact:

Q3 report published with wrong Scope 1 emissions (18% overestimated)
Auditor flagged data quality issue
Report delayed 3 weeks for resubmission + re-audit
Team spent 40 hours investigating, re-collecting, re-calculating

This error was completely avoidable.

Why Manual Review Fails

🚨

The Problem: A human reviewer looking at 10,000 emissions records can't detect statistical outliers. They see:

Facility A: 4,500 tonnes
Facility B: 500 tonnes
Facility C: 18,000 tonnes (red flag? or larger facility?)

Is 18,000 tonnes an error or legitimate? Without context, it's impossible to tell.

This is why anomaly detection requires statistics + machine learning, not just human eyeballs.

How Anomaly Detection Works

CarbonSite uses 8 concurrent detection rules:

Rule 1: Z-Score Anomaly Detection

Flags values >3 standard deviations from facility average.

Example:

Facility historical average: 50 tonnes/day
Standard deviation: 10 tonnes/day
New record: 367 tonnes
Z-score: (367 - 50) / 10 = 31.7 (extreme outlier)
Action: Flag for manual review

Rule 2: Isolation Forest

ML algorithm that identifies isolated points in multi-dimensional space (weight, date, supplier, category).

Rule 3: Time-Series Anomaly Detection

Detects unusual patterns in temporal data:

Sudden spikes
Gradual drift (emissions increasing every day)
Cyclical breaks (should have weekly pattern, but doesn't)

Rule 4: Duplicate Detection (7-day window)

Flags invoices with same vendor + amount within 7 days.

Common fraud signal
Or genuine recurring suppliers

Rule 5: Unit Inconsistency

Flags records where extracted unit doesn't match category.

Energy data should be kWh/MWh, not tonnes
Waste should be kg/tonnes, not litres

Rule 6: Out-of-Range Detection

Flags physically impossible values:

Negative emissions
Weight 0 (but energy > 0)
Date in future

Rule 7: Missing Context

Flags records that lack required fields:

No facility specified
No emission category
No timestamp

Rule 8: Supplier Performance Drift

Flags suppliers whose data quality is declining:

Used to provide complete data, now partial
Historical accuracy 95%, now 70%

Results

After each import, CarbonSite flags:

Critical anomalies (likely errors): 2–5% of records
Warnings (possible errors): 5–10% of records
Info (FYI): 10–15% of records

Example output:

🚨 Critical: Record #4521 (Facility A, waste)
   - Weight: 367 tonnes (Z-score: 31.7, 8x baseline)
   - Reason: Extreme outlier for this facility
   - Action: Review for data entry error

⚠️ Warning: Record #4522 (Supplier B, spend)
   - Amount: $2.5M (Z-score: 2.1, unusual but plausible)
   - Reason: 20% above historical spend
   - Action: Confirm with procurement

ℹ️ Info: Record #4523 (Facility A, energy)
   - Completeness: 85% (good)
   - Note: No supplier specified

✅

Real-world accuracy: In production, anomaly detection catches 92% of data quality issues before they reach calculation engine. Manual review catches 78% (some outliers are legitimate).

Explainability Matters

When anomaly detection flags a record, it also explains why:

"This facility normally reports 40–60 tonnes/day. You submitted 367 tonnes today. That's 6× your baseline. Here's what could explain it:

Data entry error (most likely, given that 36.7 was probably intended)
Measurement error (scale malfunction?)
Legitimate spike (special event, facility redesign?)

Action: Approve the correction, or confirm it's real."

This transparency builds trust. Teams understand why the system flagged something.

Machine Learning Improvement

As more data flows through CarbonSite, the ML models improve:

Z-score thresholds adapt to your facility baselines
Isolation Forest learns your data patterns
Time-series models improve with more historical data

After 6 months of data, false positive rate drops from 8% to 2%.

Next Steps

Import Q3 data → CarbonSite anomaly detection runs automatically
Review critical flags → Confirm or correct data entry errors
Calculate clean → Use validated data for final emissions report

Detect Anomalies in Your Data

Upload your emissions data and see automatic data quality scoring.

Try Data Quality Check

Related reading: