PaperlayerView exact revision on arXiv

Source-linked AI summary

OutageDiT: A Generative Foundation Model for Power Outage Forecasting and Scenario Simulation

Yunqin Zhu, Feng Qiu, Yao Xie

arXiv:2609.01896v1cs.LGcs.AI

TL;DR

Power-outage planning needs complete uncertainty-aware trajectories despite scarce severe-event data in individual regions. OutageDiT uses cross-region pretraining with local conditioning and flow-based generation, improving forecasting and scenario quality while transferring directly to held-out Michigan; extreme-event undercoverage remains the primary limitation.

  • Problem

    Outage planning requires complete trajectories capturing uncertainty and temporal dependence, while severe outage and restoration examples are scarce and heterogeneous within individual regions.

  • Method

    OutageDiT combines cross-region pretraining with local history and covariate conditioning, using a condition encoder and shallow flow decoder to generate seven-day outage trajectories.

  • Results

    OutageDiT improves national point and probabilistic accuracy and scenario quality over reported baselines, and leads every reported zero-shot Michigan transfer metric.

  • Takeaways & Limitations

    Conditional outage simulation connects forecasting with scenario-based operational planning under uncertainty and supports direct transfer to regions excluded from training.

  • Takeaways & Limitations

    Extreme-event undercoverage remains the primary limitation.

Abstract

from arXiv · show

Power-outage planning requires scenarios before an event occurs. These scenarios must represent uncertainty in magnitude, timing, and duration while preserving temporal dependence. However, severe events are rare, and data from any single region contain few examples of extreme outage and restoration patterns. To address this challenge, we introduce OutageDiT, a foundation model for generating seven-day outage trajectories at quarter-hour resolution, trained on outage and weather records across the United States. Specifically, a condition encoder processes the historical context and known future covariates once per forecast, and a shallow flow decoder reuses the resulting horizon-aligned states to generate complete trajectories. The resulting samples support point forecasting, uncertainty quantification, and conditional event simulation within one deep generative model. Across outage forecasting benchmarks, OutageDiT improves forecast accuracy and scenario quality over strong baselines and supports zero-shot transfer to held-out regions. Together, these results position conditional outage simulation as a bridge from outage forecasting to operational planning under uncertainty.

1 Introduction

Outage planning needs complete, uncertainty-aware trajectories before outage magnitude and duration are known, but severe events are rare within individual regions. OutageDiT addresses this with cross-region pretraining and local conditioning to support forecasting and scenario generation.

  • Planning decisions such as crew staging and resource allocation require outage-count trajectories that preserve escalation, peak severity, and restoration patterns.
  • Severe outage and restoration trajectories are rare and heterogeneous, leaving single-region models with few examples of extreme events.
  • Cross-region pretraining broadens observed outage patterns while local outage history, weather, and calendar covariates preserve regional context.
  • OutageDiT is introduced as a generative foundation model for power-outage trajectories inspired by the Diffusion Transformer architecture.
  • The model targets complete seven-day trajectories and reports improved point and probabilistic accuracy, scenario dependence, uncertainty representation, and direct transfer to a held-out state.

2 Method

OutageDiT combines carry-aware count encoding, a condition encoder, and a flow-based decoder to generate nonnegative integer outage trajectories. Cached conditioning states make sampling reuse one encoder pass across Monte Carlo samples and solver steps.

  • Carry-aware decimal coordinates keep consecutive wide-range outage counts adjacent across digit carries and append normalized log magnitude for scale.
  • The condition encoder processes 14 days of history, recent outage context, and seven-day weather and calendar covariates into 168 horizon-aligned states.
  • The flow decoder jointly attends to noisy future tokens and condition states, with flow time modulating residuals through adaptive layer normalization.
  • Conditional flow matching trains trajectories between data and Gaussian noise, while an auxiliary head predicts log10(1 + y) at quarter-hour resolution.
  • At inference, cached encoder projections are reused across samples and ODE steps; hourly outputs unfold into quarter-hour velocities before integer decoding.

3 Experiments

Experiments evaluate national forecasting, scenario quality, ablations, and zero-shot Michigan transfer using shared outage benchmarks and challenge protocols. OutageDiT leads the reported forecasting, scenario, and transfer comparisons, while event-window coverage remains a limitation.

  • Experimental setup: National splits use county-level U.S. windows outside Michigan, with validation in January–February 2025 and testing in March–May 2025.
  • Experimental setup: National evaluation reports MSE and weighted quantile loss in log10(1 + y) space, with lower values indicating better performance.
  • National forecasting: OutageDiT performs best on every national split and metric, with gains maintained across normal and event windows.
  • National forecasting: Future conditioning has the largest ablation effect, while history, recent context, digit coordinates, auxiliary log-count loss, and uniform digit noise also contribute.
  • Scenario simulation: OutageDiT achieves the lowest variogram score and highest scenario coverage with interval width comparable to competing methods.
  • Scenario simulation: Event-window undercoverage identifies extreme-event calibration as the primary scenario-simulation limitation.
  • Geographic transfer: Without Michigan training, OutageDiT leads every reported transfer metric across average, normal, and tail errors, nonzero detection, and interval quality.
  • Figures: Figure 1 depicts restoration forecasting in St. Louis and event-onset forecasting in Douglas, while Figure 2 displays 64 trajectories over 24 hours and joint MMD.

4 Conclusion

OutageDiT generates conditional seven-day outage trajectories from U.S. county records, improving forecasting accuracy and temporal-dependence preservation while transferring zero-shot to Michigan. Extreme-event undercoverage remains the primary limitation.

  • OutageDiT improves point and probabilistic forecast accuracy and better preserves scenario-level temporal dependence on national benchmarks.
  • OutageDiT outperforms the reported DMDA and IISE competition baselines across all metrics on geographically held-out Michigan.
  • Ablations attribute the gains to the full conditioning structure and count representation.
  • Extreme-event undercoverage remains the primary limitation.
  • Future work should evaluate scenarios in crew-staging and resource-allocation decisions, account for weather-forecast error, and address worst-case planning under distribution shift.

A Data and Evaluation Protocol

The dataset uses U.S. county records outside Michigan for training and reserves Michigan as a geographically held-out transfer benchmark, with seven-day targets conditioned on historical counts, weather, and calendar variables.

  • Michigan forms a geographically held-out transfer benchmark, while training counties exclude Michigan.
  • Outage records span 2018–2024, and complete weather covariates are available from January 2022.
  • Each example contains 14 days of context and a seven-day target, corresponding to 1,344 and 672 quarter-hour observations.
  • Inputs include outage and customer-count history, 84 weather variables, and eight calendar variables.
  • The conditioning path includes observed weather and calendar values over the seven-day forecast horizon.

A.2 Training and Evaluation Splits

Training windows follow the natural county–date distribution, retain different history streams according to a prescribed scheme, and augment weather conditioning with masked-weather probabilities; evaluation uses event and normal windows.

  • Each training draw selects a county and then a valid forecast origin in that county.
  • The complete 14-day history stream is independently retained for half of training examples, while the overlapping 24-hour recent stream is always retained.
  • Weather conditioning uses probabilities 0.80 for full past and future weather, 0.10 for masked future weather, and 0.10 for fully masked weather.
  • The national validation and test sets each contain 30 event and 30 normal windows drawn from specified date ranges.
  • Event windows come from ten distinct events with three forecast origins per event, and events exceed 800 outages and 4% of tracked customers continuously for at least six hours.

B Architecture and Optimization

The architecture tokenizes quarter-hour streams hourly, aligns condition streams on a shared time axis, and uses cached condition states to guide a timestep-conditioned flow decoder.

  • Tokenization and Position: Four consecutive quarter-hour observations are patchified into one hourly token, yielding 336 history, 24 recent, and 168 future tokens.
  • Architecture: The condition encoder uses four transformer blocks, while the flow decoder uses two blocks; both have width 1,024 and 16 attention heads.
  • Condition Encoder: The condition streams combine count, weather, and calendar modalities and share a signed hourly time axis with context before and forecast tokens after the origin.
  • Condition Encoder: History is read-only, while recent and future streams are mutable and the encoder returns the normalized future representation Q.
  • Flow Decoder and adaLN: The decoder uses Fourier timestep embeddings to modulate attention and feed-forward branches through layer-specific maps.
  • Flow Decoder and adaLN: Projected condition states supply cached keys and values, which remain fixed across ODE solver steps and Monte Carlo samples while future-token projections are recomputed.

B.4 Model and Training Configuration

The reported configuration combines analytic count handling, carefully specified initialization, and mixed optimization procedures for OutageDiT and its supervised baselines.

  • Model configuration: The model uses an analytic count representation with uniform coordinate noise and an auxiliary log-count prediction loss weighted by 10^-3.The representation preserves valid integer-count decoding while the auxiliary head predicts quarter-hour log10(1 + y).
  • Initialization: Linear maps use Xavier-uniform weights and zero biases, while timestep and adaptive-normalization components use separate Gaussian initialization schemes.The final modulation and output projection are initialized to zero.
  • Optimization: Training optimizes transformer hidden matrices with Muon and remaining parameters with AdamW using bfloat16 mixed precision.An exponential moving average begins at update 1,000, and validation runs every 500 updates.
  • Baseline configuration: Supervised baselines share the county/origin sampler and available covariates, with validation-frozen checkpoints evaluated on test and Michigan data.Baseline learning rates are selected by validation MSE, except Chronos-2 uses its released interface default of 10^-6.

C.1 National Metrics

National metrics evaluate point accuracy, weighted quantile performance, interval calibration, and temporal dependence for generated outage trajectories.

  • Evaluation setup: National evaluation uses M = 64 scrambled Sobol Gaussian trajectories and 20 Euler steps over T = 672 forecast positions.The evaluation defines sample-based summaries from the generated trajectories.
  • Point metrics: Case-equal mean squared error is computed in transformed space using s(y) = log10(1+y), with quantile-only models represented by their median.Sample-based models use the mean across generated trajectories for the point summary.
  • Probabilistic metrics: Weighted quantile loss evaluates quantiles Q = {0.1, 0.2, . . . , 0.9}.The metric summarizes probabilistic forecast performance across the specified quantile levels.
  • Interval metrics: Prediction intervals use empirical quantiles of the 64 trajectories in transformed space at the selected central interval level.The lower and upper bounds are defined by the empirical α/2 and 1 − α/2 quantiles.
  • Temporal dependence: Temporal dependence is measured with the variogram score of order p = 0.5.This metric complements point and interval measures by evaluating trajectory dependence.

C.2 Michigan Metrics

Michigan metrics evaluate geographically held-out forecasts using county-level error, tail-event classification, and interval-quality measures under fixed summary conventions.

  • Transfer setup: The transfer evaluation covers all held-out Michigan counties and the official 48-hour target, with OutageDiT using weather through the forecast horizon.Michigan metrics are computed in original count space, unlike the national metrics that use s(y).
  • Point metrics: County-macro RMSE measures point error across county-hour forecasts, averaged over all C × 48 county-hours.The metric is defined at horizon H using county-hour point forecasts.
  • IISE metrics: IISE evaluation uses a county-specific 95th-percentile training outage count and reports normal and tail error, nonzero F1, and interval Winkler score.A county-hour is positive for nonzero F1 when its outage count is at least one.
  • Interval metrics: The 95% interval is formed from empirical 2.5th and 97.5th percentiles of 64 trajectories, with Winkler penalties for observations outside the interval.The score adds penalties proportional to the distance beyond the lower or upper bound.
  • Reporting conventions: For Table 2(b), validation-frozen medians summarize normal RMSE, while 80th percentiles summarize average RMSE, tail RMSE, and nonzero F1.Winkler uses the full sample-based interval rather than a point summary.
Loading 2609.01896v1…