Disclosure control & data provenance

Grouping the dataset into 3 customer class groups, 1 were safe to publish and 2 were withheld under the rules below. The exact share published and the grand total are withheld here so the small suppressed group can't be recovered by subtraction; the published output carries a worst-case re-identification risk of 1% (Low).

Anomalies are flagged, not dropped; rows failing validation are quarantined as rejected. 24 reads (1.7%) lack coordinates and are excluded from spatial views.

• Minimum-count (k-anonymity) — 1 group withheld (triggered by your minimum group size k = 15).
  Sweeney, k-anonymity (2002); ESSnet/Eurostat SDC Handbook (Hundepool et al.) — minimum-frequency / threshold rule · Meets/exceeds: CA CPUC D.14-05-016 ("15/15"); IL 220 ILCS 33
• (n,k)-dominance — 1 group withheld (triggered by your top-1 dominance limit p = 15%).
  ESSnet/Eurostat SDC Handbook §4.2 (Hundepool et al.) — sensitive cells in magnitude tables
• PII handling — identifiers are excluded from the analytic path; only disclosure-controlled aggregates are reported.
  NIST SP 800-122; NISTIR 8053

Consumption profile

Weather & seasonality

E = 1353 + 4.1457·HDD  ·  base 65°F · 11 months

Benchmarking — Energy Use Intensity

Energy burden & equity

Every census tract was suppressed — no tract had at least 15 accounts with an ACS median income on record.

Measurement & verification (IPMVP Option C)

Baseline model fit on 5 months before Jul 2023; savings measured over the following 6 reporting months. Synthetic demo with no intervention — avoided energy near 0% is the method working correctly.

Differential privacy — an alternative release

An alternative to suppression: instead of hiding small groups, calibrated Laplace noise is added to each count (ε = 1.0, scale b = 1.0) so no individual's presence can be inferred while totals stay useful. Mean absolute error ±0.7 across 3 groups.

Appendix — policy & citations