Reports by Guest Contributors
The transition to low-carbon systems is changing the way power systems are operated today, mainly due to renewable energy variability and forecast uncertainty, distributed energy resources (e.g., generation, storage) close to consumption centers, electrification, a growing number of power converters, and more active interactions between different stakeholders (e.g., prosumers*1, large loads such as data centers or electrolyzers, flexibility aggregators*2). Today, these systems continue to be operated by humans, relying on simulations, predefined operating procedures, expert knowledge, and conventional model-driven tools.
In this context, AI-based decision systems enable the processing of far more possible system states and the identification of solutions that humans and conventional analytical methods cannot discover quickly enough. Furthermore, they provide predictability for planning power system operations and ensure sufficient flexibility in the system to accommodate contingency scenarios, which can, for instance, result from extreme weather events or human-made disruptions such as cyberattacks.
AI developments have focused largely on perception, prediction, and generative models. The use of AI for the real-time operation and control of critical network infrastructures, such as power systems, still requires further development. In these sectors, AI is already widely used for load and renewable energy forecasting. However, its deployment in control rooms remains limited for tasks such as congestion management*3 or voltage control, particularly where the goal is to identify effective operational decisions under complex constraints. Operations research tools are well established for these tasks, although their use may also remain limited in scope or insufficiently integrated into broader operational decision-making processes.
The opportunity for AI in power systems is not replacing human decisions with faster AI-based decisions. It is creating operational systems capable of augmenting today’s capabilities to process large volumes of real-time data, anticipate risks, explore alternatives, and support safer remedial and preventive decisions in environments whose complexity is becoming impossible to manage through traditional approaches alone.
Power grids share similarities with other critical infrastructures, such as railways, airspace management, and water systems, which could make AI technology developed for one sector replicable across sectors, creating opportunities for economies of scale. The notable common features are the following: (i) these infrastructures are operated by human operators in control or supervision rooms, where information is often fragmented across multiple systems and tools; (ii) decisions sometimes need to be made in seconds, with human operators under high levels of stress; (iii) the operating state changes continuously, and future conditions are driven by external factors, such as weather and service demand; (iv) failures can propagate across large geographical areas, leading to cascading events with service disruption; (v) physical constraints cannot be violated, and human operators remain legally and operationally responsible; (vi) training data are scarce for extreme events, which can have significant impacts on the economy and citizens. These are precisely the operating scenarios in which AI can bring value and support the better operation of these infrastructures, as well as the better use of their capacity.
As more renewable energy sources are integrated to meet decarbonization targets and new loads, such as data centers and electric vehicles, are connected to the power system, there will be pressure to immediately reinforce the grid, which can require significant investment.
Reliable AI-based systems can optimize system operations and planning tasks, helping postpone grid reinforcements. For instance, AI-based assistants can recommend remedial actions to the human operator, such as real-time grid topology optimization*4, to solve congestion problems in the grid. This avoids costly redispatch*5 (or even renewable energy curtailment) and enables the grid to operate closer to physical limits without sacrificing safety. Moreover, AI can enable earlier detection of insecure conditions and recommend preventive actions, such as redispatch in the market, to mitigate forecasted congestion or voltage problems. For planning horizons ranging from months to years, grid hosting capacity for new data centers and ultra-fast electric vehicle charging stations can be assessed dynamically. AI can accelerate the evaluation of millions of operating scenarios and contingencies, supporting the development of hosting capacity maps that account for rigorous power system reliability and resilience requirements. This approach can improve the use of existing grid capacity and define operating envelopes within which large loads and power plants must operate.
Another example is AI-based dynamic line rating, which estimates the maximum current a power line can safely carry under changing weather and operating conditions, enabling better use of existing infrastructure capacity. AI can also inform asset maintenance and replacement decisions via automated fault detection and classification, as well as estimates of remaining useful life. These capabilities can help extend asset lifetimes and defer capital expenditure (CAPEX) while reducing operating expenditure (OPEX) through more targeted and timely maintenance.
The availability of experienced power system engineers is struggling to keep pace with the growing demands of the energy transition, while digital technologies increasingly compete for the attention of new engineering graduates. In this context, AI, particularly generative AI, can help preserve, transfer, and scale access to operational knowledge across engineering teams and successive generations of engineers.
Overall, in critical infrastructures, AI’s greatest economic contribution may lie in safely increasing the usable capacity of existing physical assets backed by billions of euros in investment. Realizing this potential requires faster product development, reliable AI systems, and sustained attention to user adoption, which is a major bottleneck that calls for human factors to be considered throughout the design and deployment process. These challenges are discussed in the sections below.
In critical infrastructures, high AI model accuracy alone does not ensure operational reliability. Reliability also depends on the AI system’s ability to recognize situations beyond its training distribution, respect physical and operational constraints, and enable human operators to intervene or take over control when necessary. The European project AI4REALNET*6 is addressing a gap in the state-of-the-art by establishing links between AI research and human factors, work psychology, and organizational design, and between sectors that share the same underlying problems but develop separate solutions.
Among the functionalities and properties needed for reliable AI-based decision systems, AI4REALNET is developing the following.
(1) AI must produce technically sound recommendations and quantify its uncertainty. In other words, accuracy is important when providing guidance to humans, but the AI model’s ability to assess its own uncertainty and estimate confidence in its decisions is fundamental to establishing and sustaining trust between humans and AI. This is in line with recently developed System One models*7, such as Jev*8 and Decider*9; however, AI overconfidence remains a potential problem and requires well-calibrated models.
(2) AI and humans must interact through well-designed decision processes, including escalation and fallback mechanisms. This means that systems engineering, which industry has successfully applied for decades, is still needed to design compound systems integrating humans, AI, and conventional model-driven tools. New human-machine interfaces (HMIs) are also needed to unify and synthesize the data and information presented to operators, enabling more dynamic interaction between humans and AI through the interface by drawing on the capabilities of frontier multimodal large language models.
(3) The resulting actions must preserve physical security, reliability, and continuity of service. A promising direction for AI technology development is to combine data-driven learning with explicit engineering knowledge and physics. Part of the operational knowledge can be represented through rules, physical constraints, or symbolic relationships that engineers can inspect and validate, alongside numerical models whose behavior may be harder to interpret. This can improve security and reliability.
These three functions (self-awareness, human-AI system design, and hybridization of data and knowledge) are fundamental to accelerating user adoption of the technology and its integration with legacy systems.
In addition to the technical solutions discussed above, new approaches and methodologies are required to move AI from laboratory and proof-of-concept settings to operational systems in high-stakes scenarios such as power systems.
CIGRE Technical Brochure 946*10 provides a practical view of this transition*11. Drawing on a survey of 109 participants (58 of which were system operators) from 26 countries and experience from ongoing AI projects, it identifies barriers such as insufficient data histories, inadequate data pipelines, difficulties validating models with confidential operational data, and differences between prototype and production environments. It also highlights the need for AI expertise, internal collaboration, and training across the organization. These findings suggest that power system operators need to plan for the work that follows a successful prototype: integrating the model into existing tools, maintaining its data inputs, and monitoring its performance during operation. This work requires sustained resources and should be included in the business case for deployment.
Three examples of good practice are described below.
The first example is the AI4REALNET multidisciplinary collaborative approach*12, in which AI scientists, social scientists, critical infrastructure operators, and technology providers work together to fully optimize the socio-technical system associated with specific use cases. They define the interaction patterns and requirements between humans and AI and develop AI-friendly digital environments that emulate the real operating conditions of these infrastructures, including HMIs. These environments create playgrounds for testing AI prototypes and conducting human-AI experimentation that leads to enhanced, co-created products that capture human requirements and factors and can be easily integrated into operational environments. This demonstrates the potential of using digital environments to test AI products with humans (end-users), use operational interfaces to evaluate reliability, workload, and decision quality, and accelerate product development, with generative AI playing an important role in this acceleration. This would build on established practice in air traffic management*13, where controllers participate in realistic simulations to evaluate new systems and human factors. Generative AI could accelerate prototyping and adaptation based on operator feedback, with changes reviewed and tested before deployment. This would create a collaborative development process in which AI serves as both the product and a tool for improving it.
The second example is Testing and Experimentation Facilities (TEFs), an EU concept that combines physical and virtual facilities where technology providers can primarily receive technical support to test their latest AI-based software and hardware technologies (including AI-powered robotics) in real-world environments. These are infrastructures co-owned and co-operated by academia and industry at the national and European levels that offer various services for AI developers, such as access to realistic operational data, real and virtual assets for testing, independent benchmarking, conformity assessment and certification, and high-performance computing resources. These infrastructures are particularly appealing to startups and small and medium-sized enterprises, which have greater difficulty than large enterprises in accessing data and testing facilities to validate their products and build business cases to attract clients.
The third example concerns how open-source software initiatives can enable developers and end users to collaborate and advance AI technology. Two initiatives that demonstrate this potential are GridFM*14 and AI.grids*15, which aim to develop foundation models for power grids*16, capable of performing multiple grid analytics functions using a single pre-trained model. Both were open-source from the beginning and produced first-of-their-kind foundation models, such as GENCO*17 in GridFM and MxGPS*18 in AI.grids, through collaboration between academia and industry. Industry partners are now testing these models, developed as a common good, using their own private data and grids. Moreover, AI.grids also helps address data fragmentation across Europe and supports the discovery and sharing of power grid data among system operators. This shows that open-source can be an effective mechanism for industry-academia collaboration today, augmented by “vibe coding” platforms*19 that enable the building of end-to-end modular prototypes for fast industrialization.
The future of critical infrastructure is unlikely to be either human-operated or AI-operated. It will be operated by systems in which human expertise, physical models, and machine intelligence continuously complement one another. In the future, we will have compound systems in which AI, humans, systems, and assets are optimized together, delivering full value to the infrastructure owner and operator. Increasing autonomy should be treated not as a technology milestone but as something that must be earned through evidence of reliability. For instance, shortages of air traffic controllers and power system dispatchers will push toward higher levels of automation. This brings the known risks of deskilling, loss of vigilance, and overreliance on automation.
For power systems, safe and reliable AI, as required, for instance, by the European Union AI Act, is not a constraint on innovation. It is the mechanism through which AI innovation becomes deployable and ultimately valuable. The cost associated with these developments has a direct economic return from better use and planning of the existing power grid and its assets. It can even support a paradigm shift, such as moving from a system traditionally operated in a centralized manner to distributed intelligence, benefiting from advances in agentic AI and frugal methods*20 to increase the power system’s technical and economic efficiency.
The author acknowledges that the ideas presented here were developed within the activities of the AI4REALNET project (Grant Agreement No. 101119527) and the AI-EFFECT project (Grant Agreement No. 101172952), both funded under the European Union’s Horizon Europe Research and Innovation program. However, the views and opinions expressed are those of the author only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them. The author would also like to thank the members of the GridFM and AI.grids consortia.
Ricardo J. Bessa
Associate Professor and Senior Researcher, INESC TEC, Faculty of Engineering, University of Porto
Ricardo J. Bessa, IEEE Fellow (Class of 2026), earned his 5-year degree in Electrical and Computer Engineering (2006), M.Sc. in Data Analysis (2008), and Ph.D. in Sustainable Energy Systems (2013) from the University of Porto. He coordinates the Center for Power and Energy Systems at INESC TEC. His research spans energy forecasting, computational intelligence, and smart grids. He led European projects such as AI4REALNET as Project Coordinator and contributed to FP6 ANEMOS.plus, H2020 Smart4RES, H2020 InteGrid, HE ENERSHARE, and HE AI-EFFECT. He received the ESIG (Energy Systems Integration Group) Excellence Award (2022) for contributions to advances in renewable energy forecasting.
Author’s Introduction
Ricardo J. Bessa
Associate Professor and Senior Researcher,
INESC TEC,
Faculty of Engineering, University of Porto
We provide you with the latest information on HRI‘s periodicals, such as our journal and economic forecasts, as well as reports, interviews, columns, and other information based on our research activities.
Hitachi Research Institute welcomes questions, consultations, and inquiries related to articles published in the "Hitachi Souken" Journal through our contact form.
