AI in energy: power markets, grids, and weather forecasting

Why the power sector is a natural fit for AI

Electricity has to be balanced continuously, prices clear in short intervals, and both supply and demand respond to weather. That combination produces large volumes of time-series data and frequent decisions with measurable outcomes, which is why AI is used across forecasting, bidding, asset modeling, and system planning. The U.S. Energy Information Administration publishes wholesale price and load data that shows how quickly conditions change across regions.

The same features make mistakes expensive. A weak forecast can lead to poor bids, a battery charged at the wrong time, or a network investment in the wrong place. Useful evaluations therefore start with how a model's output becomes an action and who is accountable for that action, rather than with model architecture.

Weather and power-price forecasting

Weather drives load, wind and solar output, and outage risk. Jua develops atmospheric foundation models and an agent called Athena aimed at weather-sensitive energy trading and utility workflows; its energy-trading product covers wind, solar, and power-output signals delivered through dashboards, APIs, and SDKs. TotalEnergies' accelerator program has identified Jua as an AI provider for weather-dependent energy trading. Coverage and resolution vary by product and geography, so buyers should test forecasts for the specific regions and horizons they trade.

Forecast evaluation should use a held-out period that includes extreme events, not only average conditions. Compare against the forecast already in use, measure error at the time a decision was actually made, and record how often the model's output was overridden and why.

Storage dispatch and market participation

Battery storage earns value by charging and discharging at the right times, which makes it a direct application of price forecasting and optimization. Gridmatic applies machine-learning forecasts and automated optimization to wholesale market participation, battery dispatch, commercial energy supply, renewable procurement, and flexible load. Energy Vault documents Gridmatic's ERCOT offtake and AI forecasting role for the Cross Trails battery project.

Because a company like Gridmatic can act as a market participant rather than only a software vendor, asset owners should separate forecast performance from bidding logic, contract terms, risk limits, telemetry, and their own control rights. Large generators also operate in these markets directly. Constellation operates a nuclear-led generation fleet and participates in wholesale markets and retail supply, and NextEra Energy Resources develops and operates wind, solar, storage, and other generation alongside the regulated Florida Power & Light utility.

Grid assets, planning, and system modeling

Utilities use AI and simulation to decide where to inspect, trim vegetation, harden lines, and invest. Neara combines LiDAR, GIS, imagery, and utility asset records in engineering-grade network models that simulate behavior under wind, ice, fire, flood, loading, and vegetation scenarios; Ausgrid documents an approved Neara portal and validation workflow for network design. Duke Energy's profile shows why this matters at scale: it operates regulated electric utilities across six states and is investing in grid modernization and storage.

Long-term planning is a different problem. Bayesian Energy develops Convexity, a modeling platform for capacity expansion, power markets, storage, transmission, and scenario analysis, with natural-language model editing alongside Python; Rockefeller Foundation research names the company as its modeling partner for a study of nuclear deployment in emerging economies. Here the output informs policy and investment rather than minute-by-minute operations, so transparency of assumptions matters more than speed.

Data, governance, and the limits of automation

Every use case above depends on the same foundations: timely telemetry from assets, clean historical prices and load, weather inputs with known provenance, and a record of what the model recommended and what was actually done. Without that record it is impossible to tell whether a good result came from the model, the market, or luck. Teams should store forecasts as they were issued, not as later revised, and keep the version of the model that produced each one.

Automation also needs explicit limits. A dispatch optimizer should operate inside position, state-of-charge, and price limits set by a risk owner; a grid model should feed a human-reviewed investment plan rather than approve capital itself. These limits are not a sign of distrust in the model. They are what allows an organization to expand the model's role safely once it has a track record.

Finally, separate the vendor's role from the outcome. A software provider sells a tool, a power marketer takes market positions, and an advisory firm builds a study. Each carries different incentives and liabilities, and contracts should say clearly who bears the cost of a forecast error.

Using this map

Companies are listed alphabetically and appear because their BTU Graph profiles carry public evidence for the work described. Inclusion is not an endorsement or a ranking, and private deployments are not visible to BTU Graph. Confirm current capabilities, coverage, and commercial terms directly with each provider.

A sound evaluation names the decision, the data feeding it, the benchmark it must beat, and the person who can stop it. Test with historical periods that include stress events, keep weather, price, and load inputs versioned so results can be reproduced, and decide in advance which outcomes would end the pilot.