Anomaly Detection
Real records flagged by the ingestion-layer detector (statistical z-score + IsolationForest ML, per source). Suspicious records are tagged here, not dropped, so downstream systems can distinguish a real operational event from a data-quality issue.
Total Data Points
NSW1, last 7 days
Anomalies Detected
in selected period
High Severity
in selected period
Normal Points
in selected period
Detection Model
No detections yet
Anomaly Overview
Loading real demand data…
Statistical
Per-batch z-score against each column's own mean/std within the current fetch.
ML (IsolationForest)
Per-source IsolationForest over numeric columns. Catches multivariate outliers the statistical check misses.
Hybrid
Both statistical AND ML flagged the same row. Highest confidence.
Anomalies detected per day (last 7, real)
Loading…
Detected Anomalies
| Detected | Region | Severity | Method | Score | Status |
|---|---|---|---|---|---|
| No anomalies match the current filters. | |||||
Select a row to see its detail.
How the detector works
1. Statistical layer
Each ingested record is checked for statistical (z-score) outliers against its own column's mean/std within the current batch. Failures land on this page tagged statistical.
2. ML layer
A per-source IsolationForest (scikit-learn) scores each record across its own numeric columns. Records above the anomaly threshold are flagged, tagged ml.
3. Hybrid agreement
When both the statistical check and the ML model flag the same row, it's marked hybrid — the highest-confidence class.
4. Disposition
Flagged records are never dropped — they are tagged with anomaly metadata and continue through the pipeline. Admins acknowledge, mark resolved, or mark as false-positive (e.g. a real, planned operational event).
A third, rule-based layer (out-of-range bounds, missing-value flagging) ran until 2026-08-12, when it was retired — it accounted for the large majority of flagged rows, most of them structurally expected rather than anomalous. Rows it already flagged are still real and still shown here, tagged rule — nothing new lands in that bucket going forward.