How Can AI Make Power Grids Greener and More Resilient?

Artificial intelligences (AI) consume energy, yet they can also help cut emissions in the power sector. The authors separate two approaches: “green AI,” which makes environmental costs transparent, and “frugal AI,” which designs algorithms to be efficient from the start. To make these worlds talk, they propose standardized carbon metrics that translate computing choices into real grid impacts.

These metrics are then embedded in two planning schemes. The first is an iterative loop: grid sensors send load and emission signals, the AI responds in real time to optimise consumption. The second is a long‑term optimisation model that learns flexible‑load behaviours (like programmable appliances) and folds them into generation scheduling.

The real breakthrough is the ability to switch from an “emissions‑first” to a “stability‑first” priority when the grid is stressed, such as during a blackout. In those moments, AI signals recalibrate to prioritise stability, delivering targeted relief and speeding up restoration. In practice, AI becomes a “coordinator” that balances sustainability and reliability, turning the grid into a smarter, more resilient system.

The proposed carbon metrics rely on two key components: the data centre's carbon intensity (gCO₂/kWh) and the temporal profile of computational load. By integrating these data with local energy mix emission factors, a "carbon footprint per AI operation" can be calculated in real time. The signal‑response iterative loop uses predictive control algorithms based on hidden Markov models to anticipate demand and supply fluctuations, smoothing load peaks. For long‑term planning, a mixed‑integer linear programming (MILP) formulation incorporates flexible‑load variables, network capacity constraints, and carbon‑weighted costs. The shift from "emissions‑first" to "stability‑first" is governed by a dynamic priority function defined by a grid stress index (RSI) that combines frequency, voltage, and reserve margins. When RSI exceeds a critical threshold, the emission weight in the optimisation model is reduced, favoring solutions that preserve electrical stability even at temporary higher emissions. This hybrid approach ensures AI acts as an active player in sustainable and resilient energy infrastructure management.