Why this event matters
The May 21, 2022 derecho is an ideal benchmark for utility intelligence because the event had clear meteorological drivers, major vegetation interaction, a long damage path, large customer impact, and high operational consequence. Hydro-Quebec later reported that the storm crossed more than 300 km of Quebec with gusts up to 150 km/h, caused 11,254 outages, and affected 554,649 customers at the evening peak. Environment and Climate Change Canada described the broader Ontario-Quebec derecho as a billion-dollar storm that travelled roughly 1,000 km from Sarnia toward Quebec City in about nine hours.
The no-future-information rule
A trustworthy reconstruction must not use outcome data to make the prediction look better. In a GeoGridIQ backtest, the model clock would be frozen before the derecho reached Quebec. Inputs would include only historical weather forecasts and observations available at that time, archived outage history, seasonal vegetation indicators such as NDVI, known critical infrastructure locations, and regional reliability patterns. Post-event outage counts, restoration statistics, damage reports, and vegetation-cause percentages would be held back until the validation step.
What was knowable before the storm
A 24-hour pre-event view would focus on heat, humidity, severe thunderstorm potential, regional wind-risk guidance, vegetation density, and historical outage vulnerability. A 12-hour view would tighten the storm corridor and emphasize southern and western Quebec exposure. A six-hour view would increase confidence as the convective line organized. A one-hour view would shift from broad preparedness to operational monitoring, critical-asset review, and crew-staging readiness. The key point is that the model would score risk before customer interruptions appeared, not after the outage map had already lit up.
Historical weather signal
In the simulation, weather would be the strongest short-term driver. Derecho risk is not just ordinary wind; it is fast-moving convective wind, often paired with lightning, heavy rain, pressure changes, and rapid spatial escalation. GeoGridIQ would classify wind gust forecasts above 90 km/h as severe weather-driven outage risk. The public after-event record confirms why this threshold matters: Hydro-Quebec reported Quebec gusts up to 150 km/h, while federal weather reporting described damaging straight-line winds, tornadoes in Ontario, and an unusually wide damage corridor.
Vegetation exposure
Vegetation would be the second major driver. May foliage increases wind loading because trees are leafed out, and dense corridors can turn severe wind into broken limbs, downed lines, blocked roads, and difficult restoration. A historical simulation would use NDVI and vegetation-density context to flag exposed corridors before the storm. This matters because Hydro-Quebec later reported that vegetation caused 90% of the damage leading to outages. A useful model would not know that future statistic, but it would know that dense vegetation under derecho-level wind is a serious outage multiplier.
Historical outage vulnerability
The backtest would also rank regions by prior outage frequency, recurring vulnerability, and customer-impact history. Historical outage density is not a guarantee that the same locations will fail again, but it gives the model evidence about where weather stress has previously translated into operational disruption. For this event, the model would pay particular attention to regions later reported as heavily affected: Outaouais, Laurentides, Lanaudiere, Mauricie, and Capitale-Nationale.
Representative GeoGridIQ simulation output
A representative GeoGridIQ historical-simulation run would likely have elevated the event well before the peak outage count. In this demo output, Outaouais scores 88% probability with Severe predictive risk and 82/100 confidence; Laurentides scores 86% probability with Severe risk and 84/100 confidence; Lanaudiere scores 79% probability with High risk and 78/100 confidence; Mauricie scores 73% probability with High risk and 76/100 confidence; Capitale-Nationale scores 68% probability with Elevated-to-High risk and 72/100 confidence. These figures are presented as demonstration outputs for the model behavior pattern, not as a claim that GeoGridIQ was running operationally in May 2022.
Explainable AI drivers
The model explanation would be more important than the probability alone. For the highest-risk regions, the expected driver mix would be wind gust forecast plus 42%, vegetation density plus 24%, historical outage frequency plus 18%, and infrastructure exposure plus 16%. In an operator-facing briefing, GeoGridIQ would translate those signals into plain language: weather-driven predictive risk is severe, vegetation exposure is a risk multiplier, historical vulnerability increases confidence, and critical infrastructure should be reviewed before the storm line crosses the service territory.
Critical infrastructure exposure
A utility-ready prediction cannot stop at probability. GeoGridIQ would overlay forecast risk with hospitals, substations, telecom hubs, emergency services, water facilities, and transport corridors. The goal would be to identify assets inside or near likely outage zones before restoration pressure begins. For a derecho scenario, that means pre-reviewing backup-power assumptions, access routes, priority restoration dependencies, and regions where tree debris could slow crews.
What actually happened
Only after the simulated forecast is stored should the actual event outcome be revealed. Hydro-Quebec later reported more than 550,000 customers without power at the peak, 11,254 outages in total, over 2,000 employees mobilized for 11 days, 1,125 poles replaced, more than 400 transformers replaced, and 40 km of electric cable installed. The hardest-hit Quebec regions included Outaouais, Laurentides, Lanaudiere, Mauricie, and Capitale-Nationale. That post-event evidence is validation data, not prediction input.
Prediction versus reality
A validation pass would compare predicted high-risk regions against observed outage impact. For this demo, the important result is not perfect accuracy; it is whether the system would have identified the broad risk corridor early enough to support preparedness. A useful backtest would measure coverage, false positives, false negatives, lead time, and confidence. If the model flagged Outaouais, Laurentides, Lanaudiere, Mauricie, and Capitale-Nationale before peak impact, that would represent strong regional coverage. Misses would still matter and should be reviewed for missing weather inputs, local asset condition, or insufficient vegetation context.
Operational value
The operational value of this kind of prediction is lead time. A utility could use the forecast to pre-position crews, brief leadership, review critical facilities, prepare customer communications, coordinate with emergency management, and monitor vegetation-heavy corridors. The model should not tell operators that an outage is guaranteed. It should say where probability, confidence, and consequence are high enough to justify preparation before the first large outage clusters appear.
Public source notes
The after-event facts in this public case study are based on Hydro-Quebec's June 14, 2022 derecho recap, Hydro-Quebec's storm outage update page, and Environment and Climate Change Canada's Top 10 Weather Stories of 2022. Those sources are used only after the simulated prediction output is described. In a production historical-simulation workflow, each forecast run would store source metadata, snapshot timestamps, feature availability, prediction output, confidence, and validation results.
Lessons from the reconstruction
The derecho reconstruction shows why trust matters more than alarmism. GeoGridIQ should raise severity when the pre-event evidence supports it, explain the drivers, distinguish predictive risk from live operational pressure, and then measure the result after the event. The model would likely have performed well on regional weather and vegetation risk, but it would still need accurate local asset data, fresh weather inputs, and calibrated historical samples to reduce misses and false positives. That is exactly why GeoGridIQ pairs prediction with validation.