What the 24% hourly probability actually measures
The headline number answers a narrow question: given a particular combination of cold exposure and system demand, what is the chance that a new winter-driven contingency event begins in a given hour? It is not the chance that a generator is unavailable all winter, and it is not a system-level loss-of-load probability. The paper estimates it from an hourly New York dataset of 20,904 records that mixes event and non-event hours, so the denominator is generator-hours of exposure rather than events.
That distinction matters because the probability is small in ordinary cold. Below a cold exposure index (CEI) of about −0.34, predicted hourly probabilities are effectively zero regardless of demand. Near the middle of the observed range (CEI ≈ 0.96, normalized demand ≈ 0.62) the estimate is about 1%. The 24% figure appears only at the extreme corner of observed conditions, CEI = 3.19 and normalized demand = 0.86, with a 95% credible interval of 14% to 36%.
The authors note that low hourly probabilities reflect how rare winter-driven contingencies are, not a weak relationship with weather. Over an extended cold snap the hourly values accumulate, which is why the framework is built for hour-by-hour simulation rather than a single seasonal number.
Cold exposure is persistence, not one cold night
The cold exposure index is the paper's way of turning public weather station data into a state-level measure of how long and how severely temperatures stay below freezing. It is constructed to capture both persistence and severity of subfreezing conditions, rather than a single daily minimum. This is the variable that lets the model connect weather to generator records that carry no plant location.
Demand enters as normalized system demand, a zonal measure of electricity load rather than the loading of any individual unit. In the first stage, the estimated demand coefficient is larger than the cold coefficient: 5.30 (95% credible interval 2.44–8.09) versus 1.17 (0.93–1.39). The authors read this as system loading amplifying cold-induced risk. Holding CEI at its 75th-percentile value and raising normalized demand from 0.39 to 0.85 lifts predicted hourly probability from 1% to 7%.
The two inputs are only moderately correlated in the hourly data (Pearson r = 0.425, variance inflation factor 1.22), so the model is not simply relabeling cold as demand. At the event level the association is weaker still, with correlations of −0.190 and −0.292 across the two parts of the capacity model.
Three stages answer three different questions
The framework is sequential, and each stage has its own outcome and its own denominator. Stage 1 estimates whether a contingency event starts in an hour. Stage 2 asks, conditional on an event occurring, how much net available capacity remains, expressed as a fraction of nameplate rating. Stage 3 estimates how long the event lasts. Mixing these up is the easiest way to misread the paper, because the 24% belongs to Stage 1 while the 13% belongs to Stage 2.
Stage 2 is itself two-part: a logistic component estimates the probability that an event leaves non-zero capacity, meaning a partial derating rather than a full outage, and a linear component estimates expected retained capacity given that partial derating. The two are combined by pairing posterior draws across a covariate grid. Stage 3 splits again, using a lognormal regression on capacity loss for partial derates and an intercept-only lognormal model for full outages.
All stages are fitted in NumPyro with Markov Chain Monte Carlo sampling, two chains, 1,000 warm-up iterations and 2,000 retained draws per chain. Reported intervals are posterior means with 2.5th and 97.5th percentiles, so they describe parameter uncertainty under the specified models.
Severe cold shifts events toward full outage
Conditional on an event, retained capacity falls sharply as stress rises. Under mild conditions (CEI = −2.3, normalized demand = 0.50) the posterior mean normalized net available capacity is 0.66, meaning affected generators keep about two-thirds of nameplate capacity on average. At the most severe observed conditions it drops to 0.13, with a credible interval of 0.03 to 0.30.
The lower bound approaching zero across much of the high-cold, high-demand region reflects substantial probability mass on full outage. At the parameter level, the cold coefficient in the outage-versus-derate classification is clearly negative (mean −0.577, interval −0.998 to −0.175), so greater cold exposure pushes events toward losing all capacity. The demand coefficient is also negative on average (−0.773) but its interval crosses zero (−2.314 to 0.691), so the data do not pin down demand's role in that classification.
This is a statement about the 153 contingency events recorded in New York, not about every gas unit in the state. The authors also caution that the credible intervals describe uncertainty in the conditional mean, so individual events scatter more widely than the plotted bounds.
Full outages last far longer than partial derates
Duration is modeled separately for the two event types, and the gap is large. For the 73 partial derating events, predicted duration rises from 2.4 hours (95% credible interval 1.4–4.1) at mild derates with capacity loss of 0.05 to 7.1 hours (2.6–18.3) at the most severe derates with capacity loss of 0.83. Intervals widen at higher capacity loss because data are sparse and durations more variable.
For the 80 full outage events, where capacity loss is fixed at unity, the intercept-only model gives a typical duration of 12.7 hours (8.0–21.2). The authors describe this as substantially longer and more variable than partial derating events.
Out-of-fold checks show where this holds and where it does not. For full outages the predicted mean is close to observed (12.64 versus 12.23 hours), with 95.0% interval coverage across 80 records. For partial derates the highest-loss group is overpredicted (5.05 versus 3.04 hours observed), and overall root-mean-square error of 4.09 hours is close to the 3.94-hour training-mean baseline, so the capacity-loss relationship adds little predictive accuracy at that level.
Aggregated inputs limit what the estimates can support
The GADS records available to the authors carry no plant locations or unit identifiers, so contingencies cannot be matched to local weather, fuel supply or unit loading. Temperature and demand are therefore represented at state or zonal level, which smooths local variation. The reported intervals capture parameter uncertainty under the specified models but do not separately quantify the contribution of that aggregation.
Cause codes are assigned by plant operators and vary in specificity, so identifying winter-driven events required conservative filtering that may exclude events whose codes do not explicitly mention weather. The authors state the true number of cold-driven contingencies is likely higher than the cause-code structure allows them to recover, and the small samples limit precision in all three stages.
The two covariates also leave out real failure pathways: gas supply interruption and differences in plant winterization practice contribute to observed outcomes but are absent from the records, as are damage type, repair logistics and resource availability that shape duration. For decisions about an individual generator, the authors say plant-specific assessment informed by equipment condition and failure mechanisms would be more appropriate.
A regional baseline, not a transferable coefficient set
Validation is confined to New York, and the authors are explicit that this does not establish predictive performance elsewhere. Climate, generation technology, fuel supply arrangements and winterization practices differ by region and may change the estimated relationships. The three-stage structure can be applied elsewhere, but the New York coefficients should not be assumed transferable without regional calibration and validation.
The paper also does not perform an end-to-end evaluation in which event initiation, capacity and duration are simulated together across a held-out winter and scored against observed system-level capacity loss. The authors explain why: with only ten winters available, such a score would reflect sensitivity to which winters are included, which their leave-one-winter-out analysis already shows for individual coefficients.
What the framework does offer is a modular interface. At each simulation hour, Stage 1 decides whether an event is triggered given the prevailing cold exposure and demand; if so, Stage 2 supplies expected retained capacity and Stage 3 draws a duration, and the generator stays degraded for that duration before being re-evaluated. That sequential logic is what makes the output usable in hourly Monte Carlo resource adequacy simulation.
The severity corner, not the average hour, drives the risk
The paper's most decision-relevant split is between ordinary cold and the extreme corner of the observed covariate space. Across most of the observed range, hourly event probability sits near 1% or below and retained capacity stays near two-thirds of nameplate. The estimates that matter for winter resilience planning come from the combination of sustained subfreezing exposure and elevated demand, where probability rises to 24% and retained capacity falls to 13%.
Duration adds a second dimension to that split. A full outage at 12.7 hours typical duration removes capacity for far longer than a partial derate at 2.4 to 7.1 hours, so the same event count produces very different unavailable energy depending on which type dominates. The negative cold coefficient in the outage-versus-derate classification is the mechanism linking colder conditions to that longer, deeper loss.
The authors frame the contribution as the modular framework rather than predictive precision, and the validation results support that framing: full-outage duration is well calibrated, while partial-derate duration prediction barely beats a training-mean baseline. Readers weighing these numbers for planning should treat the extreme-condition probability and capacity estimates as a regional statistical baseline to be recalibrated with plant-level records, not as unit-level forecasts.
Sources used in this answer
A multi-stage probabilistic framework to estimate gas-fired generator performance during extreme winter weather
Under the most severe observed New York conditions, estimated mean hourly contingency probability reaches 24% and expected net available capacity falls to 13% of nameplate; full outages have a median duration of 12.7 hours.
