AI represents a huge opportunity for grid operators facing a rapidly changing load landscape, but there is little room for error. Moving away from trusted legacy models is risky and operators are tentative, but there is appetite for efficiency.

Across the board, the sector is experimenting. AES uses an AI-enabled Smart Operation Centre to integrate grid data for enhanced source management, while E.ON’s Intelligent Grid Platform is a smart grid technology platform that unites grid data.

Elsewhere, National Grid has partnered with Emerald AI to explore AI and managing grid flexibility, exploring the role of AI as both a load source and manager. Meanwhile, Hydro-Québec has seen success with its AI load forecasting model.

Many grids are already 'smart'. Smart grids are electricity networks that use digital communication, sensors and automation to monitor, control and optimise the generation, transmission, distribution and use of electricity in real time. However, as grid operators now teeter on the edge of widespread AI adoption for load forecasting and demand curve calculations, AI could make smart grids smarter than ever – in addition to massive potential efficiency savings.

“Load forecasting has always been an important function. AI is both a challenge and an opportunity on both sides of that equation. It is complicating load forecasting because of the demand it generates; however, we now have this opportunity to use this new tool to do better load forecasting," explains Emerald AI commercial business lead Aroon Vijaykar.

“At a minimum, AI can help to run systems more efficiently. At a maximum, it will avoid catastrophic blackouts."

Legacy models meet changing consumption patterns: AI’s moment to shine

Historically and, for the most part, currently, traditional load forecasting has involved using a mathematical model alongside years of previous load curves, tracked across relevant seasons and periods.

Sylvain Clermont, lead author of the UNECE Task Force on Digitalisation in Energy’s case study on Hydro-Québec’s AI use, explains: “We look at patterns and try to find a model that gives a curve to match, then we adjust parameters depending on the day, and fit them to the mathematical model until it looks right.

“With experience, you have a lot of historic curves, so our mathematical models are quite good for regular patterns,” he adds. “Then comes something totally out of the box – whether extreme weather or something else that you have never experienced in the past – and your model will be off.”

He points to the pandemic during which employees stayed home, industries ground to a halt and power demand changed overnight. Grid operators had no similar event – and therefore no historic curve – to work from.

It is one example from a host of new complicating factors threatening the reliability of legacy models on both the generation and demand sides, including unpredictable weather events, the AI data centre boom and the integration of renewables, which offer a less predictably stable power source. Non-integrated renewables offer their own complications, as consumers increasingly use rooftop solar panels, reshaping their reliance on the grid.

Head of network architecture and innovation at National Grid, David Adkins, tells Power Technology: “The increasing penetration of renewables introduces variability and uncertainty into grid operations, making traditional forecasting and control methods less effective. AI adoption enables real-time analysis of complex, multi-source data and supports dynamic grid management, which is crucial for integrating intermittent sources like wind and solar while maintaining stability and reliability.”

Development of AI technologies has coincided with rapid change in generation and demand, and a resultant ever-growing gap between legacy models’ load forecasts and actual demand. “The gap is growing [in matters of hundreds of megawatts]. The number of days where our traditional model is not good, are growing, but on a typical day both AI and legacy models would give you a pretty accurate forecast,” says Clermont.

The question then is not one of bad legacy models but of AI potential in grid preparedness, particularly amid swiftly changing climates and technology demands. For operators, integrating AI in smart grids enables a proactive, rather than reactive, approach.

“AI complements but does not replace grid planning and forecasting. AI-driven orchestration works best alongside grid forecasting, market signals and human operational oversight. It is not a stand-alone solution but part of a broader flexibility and resilience toolkit,” notes Adkins.

Hydro-Québec: AI load forecasting in action

Montreal-based hydropower utility Hydro-Québec is one of the largest hydroelectricity producers globally, operating more than 60 hydropower stations, and a crucial grid operator in Canada. An early adopter of AI, it began using it daily for load forecasting in 2024.

Hydro-Québec uses AI in short-term load forecasting, up-to-the-minute forecasting within a 36-hour period and hourly load forecasting between ten and 12 days in advance. This longer forecasting uses meteorologists’ daily forecasts, but the company also uses AI for up to 42 days of hourly forecasting using the “historic normal” weather data values.

Speaking to Power Technology, a spokesperson from Hydro-Québec explains that the company’s AI strategy “was not a business priority nor to be an early integrator of AI in load forecasting”. Instead, the development and integration of AI in its grid originated organically from a 2018 proof of concept using a simple neural network model to compute load forecasting for a single substation within the power grid.

The company compared the results with its legacy model forecasts, and the results were so significant that it started a research and development project.

The company is now using AI as its main model but continues to run the older legacy models alongside it. These models, the spokesperson explains, are non-linear with constraints models, based on ENLSIP (Easy Nonlinear Least-Squares Inequality Programme) algorithm estimations and a base of several tens of functions with hundreds of parameters, adjusted regularly.

The parallel use of AI and legacy models enables comparison to identify significant disparities, which human intervention can then address.

“When they get confident, Hydro-Québec will stop using the old model,” says Clermont. “It is a training question; the AI needs to be trained. On the first day, it is probably not that good, but after a year, it is probably better than you.”

AI uses machine learning to identify patterns and correlations hidden amid complex datasets; it continuously learns and adapts its predictions, enabling it to rapidly adjust to shifts in generation and demand. In turn, the AI’s learning enables utilities to manage supply and demand dynamically, balance storage and prevent outages.

However, it takes time to achieve this level of efficiency. Hydro-Québec conducted five years of research before putting its deep neural networks into production in October 2023 for load forecasting.

In its 2024 AI integration assessments, the company reported that during a heatwave on 22 May 2024, the oldest of its two legacy models failed to anticipate that the grid would not experience its typical load decrease. It required intervention by an operator and “significant” corrections of 1,500MW.

Meanwhile, the AI model successfully predicted the absence of the typical load decrease.

Clermont sees these unusual moments as AI’s opportunity to shine: “We are starting to see AI models be better at things that are not normal. When we have something unusual, they see it and they can model it. We are moving to AI not because the other models are bad, but because the few days that are bad are becoming worse and more frequent.”

Across 2026 and 2027, Hydro-Québec plans to achieve continuous improvement of operating AI, continued work with smart meter data and a renewable energy forecasting prototype. From 2028, it will commence a bottom-up, regional approach using AI to provide load forecasting for more than 350 substations.

National Grid and Emerald AI: load generation and self-management

Outside of load forecasting, smart technologies are already widely adopted. According to Power Technology’s parent company, GlobalData, between 70% and 75% of customers in the US had advanced metering infrastructure as of the early 2020s, while China has seen around 80% adoption in smart meters deployment. In the EU market, GlobalData tracks smart electricity meter penetration of between 80% and 90%.

However, using this wealth of smart, AI-driven insight on a grid-wide scale represents a significant challenge. Referring to its 2028 goal, Hydro-Québec’s spokesperson comments: “Having to deal with data from more than four million smart meters is another ball game.”

Extensive AI networks are risky. Emerald AI’s Vijaykar notes: “Utilities are understandably conservative about the system. Unlike the tech industry, which can afford to move fast and break things, for utilities one mistake can be catastrophic. They need to really kick the tyres on new technological solutions.”

For Emerald AI and National Grid, the priority now is feeling out what they hope to be a symbiotic relationship between AI and the grid. AI offers huge efficiency savings in load forecasting, but also represents a huge, unpredictable load source; in particular, model training is power-intensive and difficult for operators to predict, making the development of AI technologies part of the problem they are touted to solve.

However, by focusing on flexibility, power-hungry AI technologies could shape the grid and their own load according to the reserves of operators and demands of consumers. “We can receive signals from the grid and spin up into action in a very short amount of time with very limited, advanced notice from the utility and deliver a load shape, both in terms of ramp down, per cent load reduction and MW load reduced,” explains Vijaykar. “We can manage how long we are holding that load reduction, ramping back up over a defined period of time and avoiding snapback, which is not helpful to grids.”

Specifically, Emerald AI’s Emerald Conductor, which works as a smart mediator, enables observation of a data centre’s workloads. It intelligently determines which workloads are flexible, and which are high priority to customers, using this assessment for load flexing decisions, scaling up or down the number of general processing units allocated to each job.

National Grid trialled this AI-powered load management and grid flexing solution via a real-time orchestration of non-critical data centre workloads in response to grid conditions between 15 and 19 December 2025. “We will review the results of the trial before deciding on potential further rollout,” says Adkins.

AI is here to stay, and operators’ ability to harness its efficiency-saving potential will define the sufficiency of power supplies in the years ahead.

Adkins concludes: “AI integration promises enhanced forecasting accuracy, proactive grid management, reduced operational costs and improved capacity to adapt to evolving energy landscapes.

“In the future, AI is expected to facilitate autonomous grid operations, optimise energy flows, and enable seamless integration of distributed generation and storage. This will support the energy transition and ensure the grid remains robust, flexible and sustainable.”