Cite report
IEA (2026), Modernising Grids in the Age of Electricity, IEA, Paris https://www.iea.org/reports/modernising-grids-in-the-age-of-electricity, Licence: CC BY 4.0
Report options
AI-enhanced solutions
AI not only adds a new set of applications to the digital grid toolkit; it also helps make monitoring and control systems and GETs easier to target, operate and scale. AI’s main functions are to forecast, detect, diagnose, screen and prioritise, simulate and optimise. Applied to monitoring and control systems, these functions improve forecasting, data validation, anomaly detection and asset analytics. Applied to GETs, they help identify where latent capacity exists, forecast when it can be used, screen whether that headroom remains secure under contingencies, and rank feasible remedial actions.
AI in the power sector spans a spectrum from rules-based systems through machine learning to generative and, recently, agentic AI capable of planning and acting across several steps with limited prompting. Network operators have used AI for two decades. What has changed is the scale and diversity of what models can process. New architectures can handle larger operational, asset, weather, imagery and customer data streams; GPUs and cloud computing have sharply cut the cost of training and inference; and utilities now hold more digital information than in previous technology cycles. This is opening more of the grid to AI, but not all applications raise the same operational risk. As AI moves from forecasting, detection and workflow support towards automated control, the validation, explainability, cybersecurity, fallback rules and accountability requirements become more stringent.
Functional roles of AI in network planning and operations
Weather, load, supply, congestion, asset stress, outages
Detect Faults, vegetation risks, cyber anomalies, oscillations, equipment degradation
Root causes, constraints, stability issues, asset degradation causes
Connection requests, candidate GET sites, assets for inspection, reinforcement options
Contingencies, restoration strategies, climate stress, connection impacts
Dispatch, switching, remedial actions, maintenance schedules
Note: These functional categories do not necessarily reflect a hierarchy in maturity of applications nor the governance requirements.
AI already works for planning, maintenance and forecasting, and is fast becoming a default productivity tool
Lower-risk AI applications, such as forecasting, maintenance, inspection and planning tools, are scaling first because they improve decisions and workflows without directly controlling the power system. Forecasting is a partial exception, since it already feeds automated market and operational processes with limited human intervention. These applications have advanced furthest where three conditions coincide: enough historical or operational data are available to train and validate models; errors (a missed inspection cycle, a forecast error) are visible and recoverable; and the output can be checked by planners, engineers or operators before it affects the system. Visual intelligence illustrates this pattern. Utilities use computer vision and deep-learning models to process drone, satellite, LIDAR and inspection imagery, helping identify vegetation encroachment, ice sleeves, asset wear, corrosion, fatigue and storm damage more quickly, and in some cases more precisely, than manual review.
Beyond network-specific applications, generative AI is also becoming a default productivity tool across utilities, as in most sectors, supporting drafting, coding, data analysis and internal knowledge management. These uses are not specific to grid operations, but they can accelerate the analytical and organisational work that surrounds them.
Early deployment of AI applications across selected network operators
Country | Operator | Deployment example | Reported outcome | |
|---|---|---|---|---|
France | RTE | Contingency screening | Around 20x faster than manual security assessment | |
India | Utilities | AI data platform to clean legacy grid data and create dynamic maps | Around 80% faster grid planning where legacy data gaps existed | |
China | China Southern Power Grid | Visual intelligence for transmission line inspection | Line length nearly doubled over a decade while inspection staff grew from around 60 to 80 | |
United Kingdom | NESO | Solar forecasting | Estimated GBP 30 million a year in reserve cost savings | |
United States | CAISO | Outage processing | Cut processing time from several hours to a few minutes | |
United States | PJM | Interconnection and planning studies | First workflow deployed in 2026, aimed at speeding up studies | |
United States | Southwest Power Pool | Interconnection studies | Targets 80% faster analysis; moving to production in 2026 | |
Notes: CAISO = California Independent System Operator; NESO = National Energy System Operator; RTE = Réseau de Transport d'Électricité.
Real-time controls are AI’s frontier, constrained by trust and auditability rather than accuracy
The adoption of AI is advancing unevenly and the reason is not model performance. In lower-risk applications, AI is already trusted because it forecasts, detects, diagnoses or prioritises information for decisions still made by planners and operators. Safety-critical AI, closer to real-time control, remains far more tentative. Here, accuracy is necessary but not sufficient: models also need to be explainable, auditable, secure against manipulation, bounded by deterministic fallback rules and embedded in clear chains of accountability before operators can rely on them for decisions with immediate system consequences.
Adoption of selected AI applications in electricity networks across operational timeframes and AI functional roles, 2026
OpenOperational security
The clearest opportunity, and the clearest test, sits in the operational security layer introduced in Chapter 1's three-layer framework. This is where the long-term potential is the greatest, because it is where the sector could move from deterministic, worst-case margins towards more risk-based and probabilistic decisions about how grid assets can be used safely. AI can support this shift by synthesising large volumes of probabilistic information, flagging risk-relevant conditions and ranking feasible remedial actions for operators but the operator retains authority and accountability throughout. Widespread deployment at this level will therefore depend on stronger data, validation, governance and trust. In the near term, AI is more likely to augment engineering judgement than replace it.
Emerging control-room and planning applications
Emerging applications sit at the frontier between planning and control-room operation. In the control room, AI can help screen contingencies, prioritise alarms, identifying topology actions and translate simulations into clearer recommendations in the control room. In planning, it can accelerate power-flow studies, scenario generation and connection assessments. None of this removes the need for engineering judgement; the near-term role is to help operators and planners examine more options, more quickly, and focus human attention on the decisions that matter most.
Market-related applications are emerging too, including price forecasting and anomaly detection in bidding and dispatch, although public evidence remains limited. Disruption-preparedness tools follow the same lower-risk pattern: by combining weather, asset, terrain and customer data, they help operators pre-position crews and plan outage response more quickly, without taking direct control-room decisions. This makes them well suited to today's evidence and trust constraints, and a possible template for higher-consequence resilience applications once validation improves.
For policy makers, the practical implication is staged adoption: decision-support tools can scale first, while automated switching or self-learning operation require validated limits, cybersecurity assurance and clear accountability before they can be trusted in safety-critical environments, requirements that are feeding the emerging field of trustworthy AI.
Deployment stage of selected AI applications by surveyed network operators, 2026
OpenGovernance and network readiness
AI also introduces risks that governance needs to address directly before applications move deeper into grid operations. These risks differ by use case. Consumer-facing automation raises concerns around bias, discrimination and poorly automated decisions. Utility-level deployment raises cybersecurity, safety and accountability risks. Market-facing tools can create risks around algorithmic collusion or distorted competition. At system level, AI also has indirect effects on grids due to rising data-centre electricity demand and environmental pressures. Frameworks such as the EU AI Act, which sets obligations where AI is classified as a "safety component", and guidance from regulators such as Ofgem in the United Kingdom point to the same conclusion: governance requirements rise as AI moves closer to safety-critical grid functions.
In both transmission and distribution, scaling depends less on the AI model than on whether the network is observable, data are reliable and outputs can be integrated into operational systems. A model can only forecast, detect, optimise or prioritise conditions that the operator can observe and act on. In distribution, this makes feeder visibility, data granularity, controllable assets and integration with DMS, ADMS and DERMS platforms especially important.
The pace of adoption depends on the enabling conditions around it: regulatory frameworks, data foundations, skills and organisational readiness, and trust, cybersecurity and interoperability. These conditions apply to the digital grid toolkit as a whole, but AI faces them in a sharper form. Models depend on data quality and continuity in ways static hardware and software do not, their outputs can be harder to explain and audit, and the resulting questions of accountability are more demanding when recommendations begin to affect operational decisions.
This operational caution should not obscure the scale of innovation underway. Grid-related AI patenting grew by more than 500% in the five years to 2022 and is now the most active area of patenting among enabling digital technologies, led by the United States and China. The research and investment momentum is therefore already substantial, even while deployment in safety-critical grid functions remains more cautious.
A more ambitious generation of AI agentic systems and foundation models is emerging but not yet operational
Emerging AI solutions (research or pilot stage)
Solution | Problem it tries to solve | Promise and challenges |
|---|---|---|
The control room receives large volumes of alarms, forecasts and constraint warning, but the choice of remedial action depends on operator judgement. As networks become more dynamic, the number of switching, redispatch and setpoint combinations can become too large to assess manually in the time available. | AI agents could screen many feasible actions and present ranked options for operator approval, helping reduce response times and make better use of existing assets. Their use in real-time operations remains constrained by the need for explainability, auditability and clear accountability, and robust safeguards against unsafe recommendations. | |
Many AI tools for grids are still developed for one task at a time, such as forecasting, asset inspection, dynamic ratings or fault diagnosis. This can make deployment costly and slow, especially when models need to be rebuilt or retrained as network conditions, data sources or operating practices change. | Foundation models could provide a common model base that is adapted to multiple grid tasks with limited fine-tuning, lowering development costs and spreading learning across applications. Their value will depend on access to high-quality and representative grid data, strong validation against physical constraints, cybersecurity protection and operator confidence in how results are produced. | |
Contingency assessment and what-if analysis mostly rely on detailed physical simulators run case by case. Screening very large numbers of disturbances, operating points and remedial actions can be computationally demanding, limiting their use close to real time. | World models could learn how the grid responds over time and make scenario exploration much faster, supporting operational planning and decision support. The core challenge is that learned dynamics approximate system behaviour rather than guaranteeing physical consistency, so extensive testing, boundary checks and integration with trusted simulators would still be needed. | |
Conventional machine-learning models can detect useful patterns in operational data but may not know why an answer follows or whether it respects grid rules. In critical operations, this creates a risk that recommendations appear accurate statistically but breach operating limits or are difficult for operators to trust. | Neuro-symbolic AI could combine pattern recognition with explicit grid rules, operating limits and physical constraints, making outputs easier to trace and keep within safe bounds. The approach is promising for trust and assurance, but integrating symbolic logic with deep learning at full network scale remains technically complex and still largely at research stage. | |
Training stronger AI models often requires pooling data from several operators, but operational data is sensitive and may be restricted by cybersecurity, privacy, commercial or sovereignty concerns. As a result, each utility may have too little data to train robust models on uncommon but important events. | Federated learning allows models to be trained across several organisations without centralising the underlying data, which could support shared learning while reducing data-sharing barriers. Its deployment still requires common data standards, strong governance, secure model exchange, careful treatment of bias across participants and agreement on how benefits and responsibilities are shared. |
Notes: Examples are illustrative and mostly reflect research or early-stage demonstrations rather than widespread utility deployment. These solutions are complementary and hybrid designs are an active research area. World models and grid foundation models, for example, describe different axes and are not mutually exclusive: the first refers to what a model learns – an internal predictive representation of a system's dynamics, enabling counterfactual roll-out – while the second refers to how a model is trained, through broad self-supervised pre-training on grid data adapted to many downstream tasks. A grid foundation model can also be a world model. Physics coupling is a separate question: in both approaches dynamics are learned from data, whereas neuro-symbolic AI encodes physical constraints and operational grid rules.