Start with a forecast contract, not a risk-coloured map
Outage prediction begins by defining exactly what is being predicted. The contract needs an event label, such as whether a qualifying interruption occurs; a geographic unit, such as a service area or grid cell; an issue time; a future valid window; and a decision the forecast is intended to support. A probability without those fields cannot be reproduced or compared fairly with an observed outcome. Research models also differ in territory, storm type, spatial resolution, and target definition, so results from one study cannot be transferred directly to a Canadian product. Sources: Taylor et al. (opens in a new tab) Taylor et al. (opens in a new tab) Yue et al. (opens in a new tab)
The unit of analysis matters operationally. A community-level model can help allocate attention among areas, but it does not identify the exact conductor, tree, or customer that will be affected. A point on a map may instead represent an aggregation or display centroid. The interface should therefore state the forecast geometry and avoid language that implies feeder- or asset-level precision when the model target is broader. NIST guidance similarly emphasizes documenting intended use, limits, and the conditions under which an AI system is expected to perform. Source: National Institute of Standards and Technology (opens in a new tab)
Freeze every input at the issue-time boundary
Forecast weather must be identified by provider, model or product, issuance time, valid time, file or object identity, and retrieval state. Vegetation evidence needs its acquisition and processing dates, spatial resolution, quality mask, and coverage. Historical outages need a cutoff and a stable label definition. Infrastructure and geographic layers need effective dates and known gaps. The Meteorological Service of Canada makes forecast, ensemble, radar, alert, and observation products available, but availability of a feed does not prove that an exact historical snapshot was retained for a specific run. Source: Environment and Climate Change Canada (opens in a new tab)
The no-future-information rule is simple: if a value was created, revised, or received after the issue-time cutoff, it cannot influence that forecast. Final storm observations, outage reports, restoration records, and revised datasets belong only in later evaluation. Leakage can also be indirect, for example when a rolling feature includes an event from the valid window or when preprocessing was fit on later samples. A feature snapshot and its lineage fingerprint make the boundary inspectable rather than assumed. Sources: Kapoor and Narayanan (opens in a new tab) United States Geological Survey (opens in a new tab)
Evidence: Conceptual Illustration
The issue-time boundary prevents hindsight leakage
Separate evidence available before issuance from observations and revisions used only after the forecast window.
-
Before cutoff
Versioned historical context
Only records effective by the issue time
-
At cutoff
As-issued forecast snapshot
Identity, issue/valid times, retrieval and feature hashes
-
Frozen run
Artifact and output
Model identity, complete prediction population, output hash
-
After window
Outcome and validation
Observed outages, revisions, restoration, metrics
Conceptual lineage diagram based on temporal-integrity guidance; it is not a live feed status display.
Figure sources: Environment and Climate Change Canada (opens in a new tab) Kapoor and Narayanan (2023-08-04) (opens in a new tab)
Training evidence and live inference are different systems
Training organizes historical examples so a model can learn associations between inputs and the declared outcome. Evaluation then tests a frozen approach on later, unseen periods or events. Live inference uses a promoted artifact and a new as-of feature snapshot; it must not refit on the event it is trying to predict. Chronological separation is especially important for weather-driven outages because storms, seasons, data providers, and reporting practices change over time. Random row splits can place closely related records on both sides of the evaluation and exaggerate apparent generalization. Sources: Taylor et al. (opens in a new tab) Kapoor and Narayanan (opens in a new tab)
Published research demonstrates that weather, radar, vegetation, land-cover, and infrastructure proxies can support outage-risk modelling in studied settings. It does not validate GeoGridIQ, guarantee transfer to a new utility territory, or prove operational benefit. A local system still needs representative labels, baselines, temporal holdouts, segment-level diagnostics, and documented uncertainty. A simpler rule or climatology can be the appropriate benchmark; complexity is justified only when it improves the declared decision under a fair comparison. Sources: Taylor et al. (opens in a new tab) Taylor et al. (opens in a new tab) Yue et al. (opens in a new tab) National Institute of Standards and Technology (opens in a new tab)
Pass the trust and availability gates before publishing
A forecast is not public merely because a serialized model file exists. The resolver must prove that the artifact is the protected registry identity for the requested region and horizon, that its feature contract matches the runtime, and that the local bytes match the recorded identity. The prediction service must then produce a current batch for the same scope and expose a valid generation time, forecast window, and freshness state. If any link fails, the result should be unavailable or clearly identified as a non-model fallback; a stale card must never be presented as low risk. Sources: National Institute of Standards and Technology (opens in a new tab) GeoGridIQ
As reviewed on July 26, 2026, GeoGridIQ's Quebec 6-, 24-, 72-, and 168-hour artifacts are suspended and unavailable because their protected temporal contract is unsafe or unproven. The British Columbia 24-hour artifact passes exact identity and local-cache verification, but no current 24-hour forecast batch was available in the runtime audit. Public availability therefore remains unverified and unavailable until generation and freshness checks pass. These states are controls, not model-performance results. Source: GeoGridIQ
Evidence: Geogridiq Test Output
A model file is only one gate
Expose the fail-closed checks between a registry artifact and a current public forecast.
-
1
Region and horizon
Requested scope matches a protected registry record
-
2
Exact artifact identity
Approved identity and local bytes match
-
3
Temporal and feature contract
No unproven or unsafe training/inference lineage
-
4
Current batch
Successful generation exists for the exact scope
-
5
Freshness
Generation and valid times meet the public policy
GeoGridIQ audit state as of July 26, 2026: Quebec horizons suspended; British Columbia artifact identity verified but current batch and freshness unavailable.
Figure source: GeoGridIQ
Read probability, evidence quality, and consequence separately
A useful output names the issue and valid times, geography, event definition, probability or risk class, model and batch status, data freshness, and the strongest associated inputs. Probability is the model's estimate for the declared event under its learned conditions. It is not a guarantee, a severity measurement, or a statement about restoration duration. Driver values explain association within the model; they do not prove that a specific tree, asset, or weather variable physically caused a future interruption. Source: National Institute of Standards and Technology (opens in a new tab)
Probability calibration asks whether events assigned similar probabilities occur at similar observed rates over an appropriate evaluation sample. That statistical idea should not be confused with an internal data completeness or evidence-quality label. A forecast can have a high model probability but weak input coverage; conversely, complete inputs can support a low probability. Interfaces should show those dimensions separately and retain an unavailable state when the model, batch, or observation contract is not satisfied. Sources: European Centre for Medium-Range Weather Forecasts (opens in a new tab) European Centre for Medium-Range Weather Forecasts (opens in a new tab)
Evidence: Conceptual Illustration
What a forecast card must say
Define each output field and prevent probability, confidence, consequence, and availability from being conflated.
Scroll horizontally or use the arrow keys to compare every column.
| Field | What it answers | What it does not prove |
|---|---|---|
| Issue and valid time | When the evidence was frozen and which future window is scored | That the batch is currently fresh |
| Geography and target | The spatial unit and qualifying outcome | Exact asset or customer failure |
| Probability / risk class | Modelled likelihood under the declared target | Certainty, impact, or restoration duration |
| Evidence quality | Coverage, freshness, missing inputs, fallback state | Statistical calibration |
| Model and batch status | Whether identity and availability gates passed | Future accuracy |
| Associated drivers | Which model inputs influenced this output | Physical causality |
Define each output field and prevent probability, confidence, consequence, and availability from being conflated.
Turn forecasts into reviewable decisions, then measure them
The operational value of a forecast depends on the decision and lead time. A broad early signal might support monitoring or a planning review; a nearer-term signal with better spatial resolution may support a more focused briefing. Operators still apply switching procedures, safety rules, field knowledge, resource constraints, and authoritative utility data. The model should not dispatch crews, declare an asset unsafe, or promise an outage without an accountable human workflow and evidence appropriate to that decision. Source: National Institute of Standards and Technology (opens in a new tab)
After the valid window closes, evaluation compares the stored forecast with the complete outcome population under predeclared temporal and spatial matching rules. Results should include misses, false alarms, denominators, base rate, calibration, lead time, and segment performance rather than one headline accuracy number. That loop supports promotion, continued monitoring, or a fail-closed pause when temporal integrity, data health, or model identity can no longer be demonstrated. Sources: National Institute of Standards and Technology (opens in a new tab) European Centre for Medium-Range Weather Forecasts (opens in a new tab) GeoGridIQ
Evidence: Conceptual Illustration
Probability and evidence quality are separate dimensions
Show why a probability cannot be interpreted without calibration, lineage, and availability context.
- Probability Estimated event likelihood
- Requires an explicit target and calibration review
- Evidence quality Input coverage and freshness
- Shows missing, stale, or fallback evidence
- Availability Artifact, batch, and freshness gates
- Can force an unavailable state
- Consequence Potential operational impact
- Assessed separately from physical likelihood
Show why a probability cannot be interpreted without calibration, lineage, and availability context.
Figure sources: European Centre for Medium-Range Weather Forecasts (opens in a new tab) European Centre for Medium-Range Weather Forecasts (opens in a new tab)
Scope and safeguards
Limitations and responsible use
- Published research supports feasibility in studied settings, not GeoGridIQ accuracy or transfer to a Canadian service territory.
- Historical outage labels, infrastructure proxies, and remote-sensing inputs may be incomplete, revised, coarse, or locally unrepresentative.
- Associated model drivers are not causal explanations of a specific failure.
- No live or trusted GeoGridIQ forecast, calibration statistic, accuracy result, lead-time result, or operational benefit is claimed.
- Operators must use authoritative utility data, safety procedures, and qualified judgment for operational decisions.
Frequently asked questions
Questions this article answers
What is outage prediction?
It is an estimate of a defined interruption outcome for a stated geography and future valid window, based only on evidence available by the issue-time cutoff.
Can an outage forecast identify the exact line or customer that will fail?
Only if that is the validated model target and the input geometry supports it. A regional or grid-cell forecast must not be presented as asset-level precision.
Is model probability the same as confidence?
No. Probability estimates event likelihood. Calibration, input completeness, model status, batch freshness, and consequence are separate concepts.
Why can a verified model artifact still be unavailable?
Public output also requires a successful current batch for the exact region and horizon plus generation and freshness checks. An artifact alone is insufficient.
Are GeoGridIQ forecasts currently available?
Not in the July 26 audit: all four Quebec horizons were suspended, and the verified British Columbia 24-hour artifact had no current forecast batch that passed availability and freshness checks.
Evidence register
Sources
Sources were reviewed on . Mutable sources are rechecked on the article review schedule.
-
F3 Environment and Climate Change Canada. Meteorological Service of Canada free data service (opens in a new tab).
-
M1 Taylor et al.. Machine learning evaluation of storm-related transmission outage factors and risk (opens in a new tab). 2023-06.
-
M2 Taylor et al.. Community power outage prediction modeling for the Eastern United States (opens in a new tab). 2023-11.
-
M4 Yue et al.. Bayesian outage prediction using radar measurement data (opens in a new tab). 2017-05-18.
-
V1 United States Geological Survey. Landsat Normalized Difference Vegetation Index (opens in a new tab).
-
A1 National Institute of Standards and Technology. AI Risk Management Framework 1.0 (opens in a new tab). 2023-01-26.
-
A3 European Centre for Medium-Range Weather Forecasts. Reliability diagram (opens in a new tab).
-
A4 European Centre for Medium-Range Weather Forecasts. Verification metrics guide (opens in a new tab).
-
A6 Kapoor and Narayanan. Leakage and reproducibility failures in machine-learning science (opens in a new tab). 2023-08-04.
-
GGI_MODEL_STATUS GeoGridIQ. Protected-artifact safety and runtime availability audit.