Electric reliability: separate outage frequency and duration
An electricity distributor can have fewer interruptions but longer restoration, or more short interruptions with the same average annual minutes. A single reliability number cannot distinguish those patterns. Research should rebuild customer interruptions and customer-minutes before comparing frequency, duration and event-filtered performance.
The deliverable is a US distribution-event ledger with matched customers served, qualifying interruptions, restoration intervals and exclusions. All events and counts below are fictional. This is a statistical operating worksheet, not a grid-safety assessment or a calculation of regulatory penalties.
Locate the metric definitions and reporting scope
Michigan MPSC’s Distribution System Reliability Metrics explains SAIDI and SAIFI under those headings and presents all-weather and major-event-excluded views. These definitions distinguish accumulated customer minutes from the number of customer interruptions.
EIA Table 11.3, Reliability Metrics Using Any Method, as read October 25, 2026, reports all-event and without-major-event measures, with SAIDI, SAIFI and CAIDI notes. Its footnotes distinguish methods and coverage. The table is not a pure IEEE-only ranking and does not provide this fictional company’s event records.
For an actual company obtain event identifiers, affected customer counts, each restoration stage, qualifying interruption rules, customers served and the reporting period. Match whether the series includes major events and supply interruptions. A different event filter can change a rate without changing what customers physically experienced.
Rebuild a complete two-event case
Assume one thousand customers served throughout a fixed year. Two separate qualifying non-momentary events occur, with no time overlap. Event A interrupts two hundred customers for sixty minutes. Event B later interrupts one hundred for 120 minutes. Fifty of B’s customers had also experienced A, after being restored between events.
Customer interruptions are 200 + 100 = 300. Unique customers affected are 200 + 100 − 50 = 250, a different quantity. The fifty repeat customers contribute two interruptions each, so replacing three hundred with the deduplicated 250 would understate frequency.
Customer-minutes are 200 × 60 + 100 × 120 = 24,000. SAIFI is interruptions divided by customers served: 300 ÷ 1,000 = 0.3. SAIDI is customer-minutes divided by customers served: 24,000 ÷ 1,000 = twenty-four minutes. CAIDI is customer-minutes divided by interruptions: 24,000 ÷ 300 = eighty minutes.
| Input or metric | Result | Denominator meaning |
|---|---|---|
| Unique affected customers | 250 | Distinct identities across events |
| Customer interruptions | 300 | Repeat qualifying events retained |
| Customer-minutes | 24,000 | Each affected duration weighted by customers |
| SAIFI | 0.3 | Interruptions per customer served |
| SAIDI | 24 minutes | Annual minutes per customer served |
| CAIDI | 80 minutes | Minutes per customer interruption |
Under the same population and definitions, SAIDI divided by SAIFI also gives CAIDI. This is an internal check, not permission to compare published ratios with different coverage or rounding. CAIDI is not average annual downtime for all customers; that is the role of SAIDI here.
Test staged restoration
For a different version of A, assume one hundred customers are restored after twenty minutes and the remaining hundred after sixty. Event A contributes 100 × 20 + 100 × 60 = 8,000 customer-minutes, rather than 12,000. Its final restoration still occurs at sixty minutes.
Keep B unchanged at 12,000 customer-minutes. The two-event total becomes twenty thousand, giving SAIDI twenty minutes, SAIFI still 0.3 and CAIDI approximately 66.67 minutes. Multiplying all two hundred customers by the last-restored duration would miss the partial recovery and overstate A’s burden by four thousand customer-minutes.
Obtain restoration stages rather than assume every customer returns at the last system milestone. Conversely, a system partially restored is not evidence that every affected customer has service. Preserve the stage counts so the total affected population closes without restoring the same customer twice.
Test frequency with the same annual minutes
For a separate yearly scenario, assume six hundred qualifying customer interruptions with an average duration of forty minutes, still across one thousand served customers. Customer-minutes remain 24,000 and SAIDI twenty-four, but SAIFI becomes 0.6 and CAIDI forty.
The original and this scenario share SAIDI yet have different frequency and restoration patterns. A claim of unchanged reliability based on SAIDI alone would omit that distinction. The worksheet should show both metrics and the underlying event distribution, rather than select whichever rate looks most favourable.
Neither scenario supplies unserved MWh. Customer count is not electrical load, and customer-minutes cannot be converted to energy without the relevant load and timing evidence. A large industrial account and a household are each a customer under a count-based measure, not necessarily equal energy exposure.
Reconcile major-event exclusion
Return to the original case and assume B qualifies for exclusion under the expressly applicable event method. The remaining A has two hundred interruptions and 12,000 customer-minutes. With the same one thousand served denominator, excluded-view SAIFI is 0.2, SAIDI twelve and CAIDI sixty.
The all-event view remains 0.3, twenty-four and eighty. B’s affected customers did not regain their lost minutes because a reporting filter excluded the event. Keep both views and the actual basis of the exclusion, rather than call the filtered improvement disappearance of customer harm.
The EIA notes explain that major-event treatment differs across methods. Do not impose a national statistical cutoff on a utility using a different method or infer a legal exemption from the label. This article does not calculate an IEEE threshold or decide whether a real event meets it.
Use the interruption verification sheet
| Field | Obtain | Check |
|---|---|---|
| Served population | Dated customer or meter count | Same year, company and defined basis |
| Events | Qualifying interruption register | Separate incidents, repeats and actual overlap |
| Durations | Restoration-stage timestamps | Customer-weighted minutes, not last finish only |
| Classification | Major-event and supply-loss treatment | Same method and filters across comparisons |
| Aggregation | Customer interruptions and minutes | Unique customers not used as interruption count |
| Context | Loads and financial records if needed | No automatic MWh, damages or penalty conversion |
When reports conflict, compare non-momentary definitions, customer populations, stage data, event exclusions and reporting years. The EIA coverage field has its own numerator and denominator; a published coverage value is not a direct count of the utility’s outage victims.
If raw events or restoration stages are missing, report the published rate and its stated scope without pretending the numerator was independently reconstructed. An unexplained difference belongs in a named reconciliation item, not an invented storm adjustment. Small differences between rounded ratios do not alone establish a reporting error.
Update after corrected events, restored-customer records, changed served counts or revised exclusions. Preserve the all-event experience alongside the chosen standard metric. The conclusion should distinguish how often customers lost service, how long interruptions lasted and what the reporting filter removed, without converting an operational rate into a universal safety or compliance conclusion.
Check the event register before publishing a comparison
Use event-customer intervals where records permit. Each interval starts at a qualifying interruption and ends at that customer’s actual restoration. In the main case, the fifty repeat customers each have two non-overlapping intervals. This preserves two interruption counts without assigning the same minute to them twice. If two incident labels cover one continuous loss of service, do not automatically create two separate customer interruptions.
The distinction can matter when control-room records split one incident into feeder and substation tasks. Match the task identifiers to the actual service interruption rather than count repair tickets as customers losing service. The appropriate classification still follows the utility’s stated reporting method; the analyst’s interval check exposes possible duplicate inputs without inventing a new universal rule.
Also reconcile the denominator independently. Suppose a revised report uses 1,200 served customers while retaining the same 300 interruptions and 24,000 customer-minutes. SAIFI becomes 0.25 and SAIDI twenty, while CAIDI remains eighty. The apparent improvement in the first two rates reflects the denominator change in this hypothetical, not a newly shortened outage. This sensitivity belongs beside the revision explanation.
A monthly series should not be averaged casually into an annual ratio when served populations differ. Rebuild the annual inputs under the stated method or use the reported annual denominator and document it. A simple average of monthly SAIFI values can answer a different question from total annual interruptions per matched customer basis.
For every comparison, attach a short scope record: reporting entity, year, customer definition, sustained-interruption treatment and event exclusions. Then identify whether the observed change came from event frequency, restoration duration, population or classification. This allows a reviewer to challenge a specific input instead of debating a single unexplained reliability ranking.
FAQ
Should repeat customers be counted only once in SAIFI?
No. Separate qualifying interruptions remain separate even when the customer is the same.
Does excluding a major event erase customers’ actual outage time?
No. It changes the reported statistical view, not the event customers experienced.