2026 Nuremberg Tech Arena - Digital Power
2026 Nuremberg Tech Arena - Digital Power
Digital Power
AIDC Power Supply Resilience — Hierarchical Prediction of Power Continuity Risk
Resilience Prediction of AIDC Power Supply under Extreme Scenario
Relevance
AI now runs on power as much as on chips. Large-scale GPU clusters operate continuously at extreme power densities, and even a few seconds of interruption can erase hours or days of training progress. As a result, AIDC power supply resilience is becoming a systemic risk for the AI industry chain.
However, this resilience is not fully controlled by the data center itself. Regardless of whether a facility adopts UPS 2N, UPS distributed redundancy, HVDC 2N, or direct mains 2N, its ultimate power source is still the grid. Extreme weather can degrade the regional distribution network, especially where overhead lines and radial topologies create vulnerable single points of failure. Once regional supply capacity is impaired, the risk propagates through incoming feeders to each AIDC site. The final impact then depends on the interaction between the external outage and the site’s internal power architecture. Under the same weather and grid-risk conditions, different configurations of incoming lines, UPS redundancy, diesel generators, and energy storage can lead to very different outage probabilities, critical-load coverage, and backup duration.
This is the core challenge: existing outage-prediction models usually stop at the regional level, estimating the probability, scope, and duration of grid failures. They rarely account for site-specific architecture, redundancy, or backup resources. As a result, they cannot answer what operators truly need to know: under this extreme-weather event, will my site stay up, how much load can I protect, and for how long?
Closing this gap requires a hierarchical, site-level resilience prediction model. Using weather, distribution-network risk, incoming-feeder conditions, and facility architecture as inputs, the model predicts a power supply risk score, critical-load coverage ratio, and estimated backup duration at two horizons: day-ahead prediction over 48 hours with 1-hour granularity, and hour-ahead prediction over 6 hours with 5-minute granularity.
Challenge 1 — Day-Ahead power-supply risk forecasting
Your first challenge is to predict, a day out, whether a site stays up. Given historical weather observations, regional power-supply risk values, and architecture configuration parameters, build a model that forecasts the site-level power-supply risk score for each of four supply architectures (IT Power 2N + UPS 2N / UPS DR / HVDC 2N / Direct Utility 2N), at 1-hour resolution over a 48-hour horizon. The risk score is a sigmoidal transduction of the regional outage proportion, with k and x₀ set by the site's incoming-feeder topology (single-feed; dual-feed common-source; dual-feed independent-source); critical-load protection ratio and expected backup duration are then derived from this score via rule-based functions. The model must capture how weather × regional vulnerability × architecture combine, and generalize across architectures, even though extreme outages remain rare, high-impact events that push naive models toward a trivial "always low-risk" prediction.
Accurate day-ahead forecasting is directly useful in two ways: for advance contingency planning, to schedule generator pre-checks, defer non-critical workloads, or pre-position backup fuel ahead of an approaching extreme-weather event; and for portfolio-level risk management, to compare how much critical load each architecture protects under the same forecast, informing site selection and capital allocation.
Accurate day-ahead forecasting is directly useful in two ways: for advance contingency planning, to schedule generator pre-checks, defer non-critical workloads, or pre-position backup fuel ahead of an approaching extreme-weather event; and for portfolio-level risk management, to compare how much critical load each architecture protects under the same forecast, informing site selection and capital allocation.
Challenge 2 — Hours-Ahead power-supply risk forecasting
Your second challenge is to predict, in the final hours, exactly when and how hard a site will be hit. Given the same inputs — weather, regional risk, architecture parameters — build a model that forecasts the site-level power-supply risk score at 5-minute resolution over a 6-hour horizon, strictly respecting temporal causality as conditions evolve in near real time. The finer granularity raises the bar: the model must track risk as it actually unfolds rather than as a coarse daily envelope, while still mapping correctly through the same topology-dependent sigmoid function and generalizing across all four architectures.
Accurate hours-ahead forecasting is directly useful in two ways: for real-time operations, to trigger load-shedding or switchover to backup power at the right moment — not too early, wasting capacity, nor too late, losing load; and for situational awareness, giving operators a live, quantified read on remaining backup duration as an extreme event progresses.
Data
Participants are provided a curated list of recommended sources across five dimensions, with API documentation and access guidance. Outage data: NaFIRS LV Faults (primary, distribution-district level, covering central-southern England and northern Scotland) and Eagle-I (US, optional reference). Weather data: Open-Meteo, with participants selecting one or more coordinates at 3–9 km spatial granularity. Site power-supply architecture configuration parameters (feeder topology, redundancy class, buffering) are provided by the organizing committee. Optional: power infrastructure via OpenStreetMap/Overpass API, and extreme weather events via the NOAA Storm Events Database. Evaluation will be performed with similar dataset.