Grid optimization and forecasting
A grid forecast pays off only when it changes a decision someone already makes: unit commitment, a line rating, a crew call-out, a planning scenario. The hard part today is not the weather model but large loads, where a forecast that takes every data center request at face value is wrong in both directions.
Part of AI for Utilities & Energy. 9 sources, checked .
- Projected 10-year growth in aggregated North American summer peak demand, 69% above the prior year's projection (NERC 2025 LTRA)
- 224 GW
- Projected 10-year growth in aggregated North American summer peak demand, 69% above the prior year's projection (NERC 2025 LTRA)
- North American Electric Reliability Corporation (NERC), 2026(opens in a new tab)
- Average electricity interruption per U.S. customer in 2024, nearly twice the annual average of the prior decade
- 11 hours
- Average electricity interruption per U.S. customer in 2024, nearly twice the annual average of the prior decade
- U.S. Energy Information Administration, 2025(opens in a new tab)
- Reduction Texas applies to requested data center load in its forecasts, based on how existing sites actually consume (NERC 2025 LTRA)
- ~50%
- Reduction Texas applies to requested data center load in its forecasts, based on how existing sites actually consume (NERC 2025 LTRA)
- North American Electric Reliability Corporation (NERC), 2026(opens in a new tab)
The problem
Demand forecasts have stopped being flat. NERC's 2025 Long-Term Reliability Assessment projects aggregated summer peak demand rising by over 224 GW in ten years, 69% higher than the prior year's 132 GW projection, with new data centers accounting for most of the increase.1 NERC finds 13 of 23 assessment areas face resource adequacy challenges over the same period.1 EIA expects the U.S. power sector to generate a record of almost 4,390 billion kWh in 2026, driven by growing commercial demand and a hot summer.3
Large loads are what make the forecast hard. NERC notes they add volatility because project timelines and commitments shift, and that in Texas new data center projections are cut by almost 50% from requested levels to match how existing sites actually consume.1 They also behave differently on the system: in February 2025 a transmission fault in the Eastern Interconnection was followed by 1,800 MW of data center load dropping off.2 A model trained on last decade's load shapes will not see either effect.
Reliability pressure is rising at the same time. U.S. customers averaged 11 hours of interruptions in 2024, nearly twice the annual average of the decade before, and three hurricanes accounted for 80% of those hours.4 FERC Order No. 881 requires transmission providers to use ambient-adjusted line ratings and RTOs to accept rating updates at least hourly, which ties transmission capacity directly to the quality of your temperature forecasts.5
DOE's 2024 AI for Energy report says operators need significantly more accurate weather forecasts as variable generation grows, and that they will be reluctant to adopt new tools without assurance those tools will not compromise reliability.6 The IEA estimates AI-based fault detection can cut outage durations by 30 to 50%.7 The change AI brings is resolution and cadence: feeder-level forecasts refreshed hourly with stated uncertainty, rather than one system number a day.
What we build
- 01
Feeder and substation load forecasting
Hour-ahead to week-ahead forecasts built from smart meter interval data, SCADA and weather ensembles, published as ranges with the main drivers shown, so operators see the uncertainty and not just a single line.
- 02
Large-load forecast adjudication
A scoring model that weights each large-load request by its contract and construction milestones and applies ramp shapes learned from sites already energized, feeding planning scenarios instead of nameplate requests.
- 03
Storm outage prediction and crew staging
Models that combine weather forecasts, vegetation, asset age and historical outage records to predict outages by district 24 to 72 hours ahead, so mutual assistance and crews are staged before the event.
- 04
Ambient-adjusted rating support
Line-segment temperature forecasts that feed hourly ratings, with every forecast, input and rating stored so your rating methodology can be shown to the transmission provider and market monitor.
- 05
Event review assistant for operators
A read-only assistant that retrieves procedures, past event reports and disturbance data to speed up post-event analysis. It has no path to control systems.
Where it goes wrong
The failure modes we design against from the first week, and the control for each.
Treating every large-load request as firm demand
NERC reports that some large-load projects slowed or failed to materialize while later-year requests kept growing.1 The control is to probability-weight requests by milestone, track realized against requested load for each project, and re-baseline the forecast at least quarterly.
Models that fail in the weather that matters
DOE lists extrapolation, predicting situations outside a model's experience, as a core AI failure mode for the grid.6 Extreme days are rare in training data. Score the model separately on peak and extreme-weather days, keep the incumbent forecast as a fallback, and publish uncertainty with every run.
Grid data leaving the security perimeter
Training sets often contain BES Cyber System Information, and CIP-011-3 requires methods to protect it, including when it is stored off-premise.8 Classify data before ingestion, encrypt and tokenize where needed, and keep models that touch it in environments your CIP program already covers.
A forecast nobody acts on
If the control room does not trust a model it will be ignored, whatever its accuracy. Run it in shadow mode beside the existing forecast for a full season, publish error by hour and by event type, and require operator sign-off before it drives commitments.
Rules that apply
NERC CIP-011-3 (Information Protection)
Primary text on nerc.com(opens in a new tab)Requires a documented program to identify BES Cyber System Information and protect it in storage, transit and use, including off-premise, which covers grid data used to train or run forecast models.
NERC CIP-013-2 (Supply Chain Risk Management)
Primary text on nerc.com(opens in a new tab)Requires a supply chain cyber security risk management plan for high and medium impact BES Cyber Systems, so AI software or vendors installed in those systems are assessed at procurement.
FERC Order No. 881 (Managing Transmission Line Ratings)
Primary text on federalregister.gov(opens in a new tab)Requires transmission providers to implement ambient-adjusted ratings and RTOs to support rating updates at least hourly, so the temperature forecasts behind those ratings must be accurate, timely and documented.
How to measure it
- Hour-ahead and day-ahead forecast error by feeder, with peak-hour error reported separately
- Forecast error on the highest-demand hours and extreme-weather days, compared with the incumbent forecast
- Realized versus forecast load for each large-load project, reviewed quarterly
- Outage prediction accuracy at 48 hours and crew pre-staging accuracy during major events
- Share of forecasts operators accept without manual override
Questions buyers ask
How accurate does an AI load forecast need to be before operators rely on it?
It needs to beat the forecast operators already use, on the hours that matter, for long enough to earn trust. Average error is not enough. Compare peak-hour and extreme-weather error against the incumbent over at least one full season in shadow mode, publish the results by hour, and move it into production only when operators agree the new forecast is better where it counts.
Can a utility use cloud AI for grid forecasting under NERC CIP?
Yes, if the data is handled inside your CIP program. CIP-011 requires you to identify BES Cyber System Information and protect it in storage, transit and use, including off-premise storage. Some inputs, such as public weather data, are not BCSI. Classify each input first, protect the sensitive ones with encryption or tokenization, and treat any vendor tool inside a high or medium impact system as a CIP-013 procurement.
How should utilities forecast data center load?
Separate the request from the expected load. Weight each project by where it stands on contracts, construction and energization, then apply ramp shapes observed at sites already running. Texas already discounts requested data center load heavily for this reason. Track every project's realized load against its forecast, and revisit the weights each quarter as more sites come online.
Sources
Every figure on this page comes from the sources below. We checked each link on .
- 1.2025 Long-Term Reliability Assessment(opens in a new tab)North American Electric Reliability Corporation (NERC), January 2026 · nerc.com
- 2.2026 State of Reliability: Overview of 2025 Bulk Power System Performance(opens in a new tab)North American Electric Reliability Corporation (NERC), June 24, 2026 · nerc.com
- 3.Short-Term Energy Outlook: Electricity, Coal, and Renewables(opens in a new tab)U.S. Energy Information Administration, November 10, 2026 · eia.gov
- 4.Hurricanes in 2024 led to the most hours without power in the United States in 10 years(opens in a new tab)U.S. Energy Information Administration, December 1, 2025 · eia.gov
- 5.Managing Transmission Line Ratings (Order No. 881)(opens in a new tab)Federal Energy Regulatory Commission, Federal Register, January 13, 2022 · federalregister.gov
- 6.AI for Energy: Opportunities for a Modern Grid and Clean Energy Economy(opens in a new tab)U.S. Department of Energy, April 2024 · energy.gov
- 7.Energy and AI: Executive summary(opens in a new tab)International Energy Agency, 2025 · iea.org
- 8.CIP-011-3: Cyber Security, Information Protection(opens in a new tab)North American Electric Reliability Corporation (NERC) · nerc.com
- 9.CIP-013-2: Cyber Security, Supply Chain Risk Management(opens in a new tab)North American Electric Reliability Corporation (NERC) · nerc.com