Decision-Point ETR Accuracy (DPETRA)

GMI-EU-034Energy & UtilitiesPilot-Ready

Definition

Measures the accuracy of the estimated time of restoration at the moment each affected customer needs to make a decision (stay/leave, seek shelter, arrange backup power), rather than accuracy measured in retrospective aggregate.

The Measurement Gap

Traditional ETR tracking compares the final predicted restoration time to actual restoration time and reports a mean absolute error across events. This produces a statistic ("our ETR was off by an average of 2.3 hours") that does not reflect what customers experienced during the outage. A utility that consistently publishes accurate six-hour ETRs at hour zero of a twelve-hour outage looks good on this metric while misleading every customer at the most critical decision point.

The specific failure point is the overconfidence direction of ETR error. If the AI system consistently produces ETRs that are shorter than actual restoration times (the most common failure mode, since models tend to assume best-case crew travel and repair times) customers with medical equipment dependencies, household members requiring temperature management, or limited resources who cannot afford hotels are making stay-home decisions based on an estimate that the model knows is optimistic but does not flag as such.

DPETRA enables the specific operational decision of whether to shift from point-estimate ETR communication to range-based ETR communication ("expect restoration between 12 and 24 hours") when model confidence at a given decision point falls below a calibrated threshold. That decision requires knowing how accurate the point estimate actually is at each decision interval, separated from the aggregate statistics that make the model look better than it performs at the moments that matter.

Without this measurement, utilities publish point-estimate ETRs because the model produces them, without knowing whether the point estimate at hour six is useful information or harmful overconfidence.

OMS with customer-level restoration tracking and AMI last-gasp/power-up signals make the customer-level ETR-to-actual comparison computationally feasible at event scale.

Formula

DPETRA(t) = 1 − (1/N) × Σ |ETR_published(i,t) − Actual_restoration(i)| / Actual_restoration(i)

Variables:

  • t: decision-point time after outage onset (measured at 0, 2, 6, and 12 hours post-outage; the 6-hour decision point is the primary metric (customers making stay/leave decisions)) Source: OMS event log
  • ETR_published(i,t): ETR estimate published to customer i at time t after outage onset (hours from outage start). Source: OMS customer notification system, customer portal, or IVR logs
  • Actual_restoration(i): actual restoration time for customer i measured from outage onset (hours). Source: OMS restoration records, AMI power-up signals
  • N: number of affected customers with published ETR at decision point t

Calculation Frequency: Per outage event; aggregated monthly and by storm type
Output: 0 to 1.0 scale. DPETRA = 1.0 is perfect accuracy at the decision point. Track by direction: overconfident ETRs (ETR shorter than actual) are more harmful than conservative ETRs. DPETRA(6) above 0.75 at the 6-hour decision point is the key threshold for customer communication quality. Below 0.60 during major storms should trigger a switch to range-based ETR communication rather than point estimates.

Proposed formula, pending validation pilots.

At a Glance

Category
Forecast Accountability, Storm Response, and Capital Planning

Decision Supported

Is the ETR system giving customers decision-quality information in real time, or is it publishing overconfident estimates that cause customers to make harmful decisions?

← Net Cycling Economic Value (NCEV)Hosting Capacity Model Accuracy Index (HCMAI) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment