Session Chair: Venkat Banunarayanan- NRECA

As artificial intelligence moves from the research into applications in grid planning and operations, the electric grid control room stands at the frontier of what could be one of the most consequential transformations in the history of the power system. This session explores what it means to place AI at the operator’s side—where split-second decisions balance supply and demand, integrate surging volumes of both firm and intermittent energy, and keep the lights on across increasingly complex and volatile networks. A key development is the emergence of grid foundation models: large, pre-trained AI systems that learn the underlying physics and behavior of the power network from vast operational data, promising a shift from narrow, task-specific tools toward versatile models that can forecast, diagnose, and reason across the full breadth of control room challenges.
We will confront the hard questions: How can AI deliver actionable insight fast enough to matter, while remaining transparent and trustworthy enough for operators to rely on when the stakes are highest? Can grid foundation models and machine learning algorithms anticipate cascading failures, optimize dispatch, and manage uncertainty in ways that surpass traditional tools, without introducing new risks around data quality, cybersecurity, and automation bias? And critically, how do we design a partnership between human expertise and machine intelligence that augments rather than replaces the seasoned judgment of the control room? Bringing together perspectives from system operators, researchers, and technology developers, this panel will chart both the immense opportunities and the real-world barriers to making AI a safe, reliable, and indispensable presence in the control room of tomorrow

Where AI Fits in Control Room Operations

Speaker: Brandon Harrington – AECI

Control room work runs along a spectrum, from normal operations through upset to emergency. As conditions worsen, time horizons collapse and the operator’s tolerance for uncertainty goes to zero. That spectrum is the lens for judging where AI fits. The biggest opportunity is before the upset, in planning, model building, outage coordination and contingency hunting, where compute is cheaper than expert hours and every answer can be vetted. Probabilistic tools should reach an operator only when wrapped in deterministic validation and a familiar HMI. For grid foundation models, the open question is what they uniquely unlock on that spectrum.

Human-Centred AI for Power System Control Rooms

Speaker: Ricardo Bessa – INESC-TEC

This talk presents the AI4REALNET framework for human-centric AI in critical infrastructures, with a particular focus on power system operation. It discusses key challenges in supporting control-room decision-making, including the rapid processing and prioritization of alarms and the need for AI models that are transparent, interpretable, and auditable.

The Control Room Operator as Forecaster and Risk Manager

Speaker: Adrian Kelly – EPRI Europe

Walk Nikola Tesla into a modern electricity control room and, after a moment of wonder at the screens, he would start nodding. The grammar of grid operations remains electromechanical even as the grid itself becomes digital. For a century, grid asset complexity, the operational technology (OT) toolkit and human comprehension advanced in lockstep. The power electronics revolution has broken that equilibrium. Inverter-based resources, HVDC and distributed energy have accelerated the grid, while monolithic EMS platforms on decade-long refresh cycles and operators facing ever-growing data volumes have fallen behind. The result is a widening comprehension gap at the moment the stakes are highest.

This talk argues that AI is the first technology capable of accelerating the two lagging pillars. That only holds if AI is deployed to subtract complexity, not add another screen. Building on EPRI’s Operational Capability Model, the presentation proposes AI-based solutions across the control room:

*Management by exception: machine-learning alarm rationalisation and anomaly detection.
*A unified operator interface: an AI copilot that collapses fragmented tools into one window.
*Probabilistic risk assessment: AI-accelerated forecasting and contingency screening that move operations beyond N-1.
*Pre-validated playbooks: AI-generated remedial actions approved by operators in advance.
*Continuous model validation: checking network models against high-speed measurements.
*AI-generated training scenarios: used within an Operational Readiness Centre.

The talk closes with the guardrails required for high-risk critical infrastructure: explainability, human authority, graceful degradation, and the emerging discipline of trust calibration. This means training operators to know when to trust AI, when to interrogate it, and when to override it. The goal is not to automate the operator away. It is to redefine the operator as a forecaster and risk manager.

The Control Room of the Future: Human Factors in AI Adoption

Speaker: Seong Lok Choi – NLR

The rapid integration of AI into operational control rooms raises a key human-factors question about when operators will willingly give the driver’s seat to an AI system and when they will retain direct command, a paradox illustrated by autonomous vehicles that are measurably safer than human drivers yet still prompt passengers to reach for the steering wheel. This preference is grounded in a cognitive distinction between safety and controllability, with delegation accepted in low-risk, repetitive, well-defined tasks where the operator can assume a passenger role, but rejected in high-stakes or unusual emergencies where direct control is demanded. In those high-stakes contexts, AI is most effective as a trusted assistant that is demonstrably accurate, procedurally faithful, transparent, and auditable, augmenting rather than replacing human judgment. We propose conditional delegation — “if it ain’t broke, don’t fix it” — in which AI takes the primary role only when human capacity is exceeded, existing processes have failed, or the domain is new, and we recommend that control-room interfaces be designed to preserve operator action while leveraging AI’s accuracy, consistency, and speed.

Speaker: Jonas Lundberg – Linköping University and Cognible AI

How can operators and AI maintain joint control when foundation models both interpret grid conditions and propose courses of action? Drawing on AI4realnet’s power grid case and experience from air, drone, rail, and vessel traffic management, this talk applies the Joint Control Framework to the design of human–AI teaming. The framework describes who perceives, decides, and acts, what they collaborate on, and how control is distributed across cognitive control levels and time horizons. This includes collaboration on situational understanding, goals, tradeoffs, plans, and specific interventions. Its Score notation makes control concepts explicit, enabling comparison between intended collaboration and observed operator and AI performance. The focus is on making AI support understandable, contestable, and timely, while enabling operators to shape goals, question assumptions, and revise proposed actions. The talk addresses coordination with existing automation, shifting control responsibilities, and opportunities for effective intervention as disturbances unfold and the available time for action contracts.

AI at the Operator’s Side: Physics-Validated Grid Intelligence and a Research-to-Operations Path for the Control Room of the Future

Speaker: Duane Verner – Argonne National Laboratory

This presentation describes how Argonne’s GridAI capabilities and findings from the Control Room of the Future Testbed (CROFT) project can help grid operators manage increasing operational complexity through faster analysis, physics-validated recommendations, and clear human decision authority. Within GridAI, LUMINA rapidly screens operating conditions and potential disruptions, while GridMind coordinates scientific tools and assembles traceable evidence for operator review. Conventional engineering models validate uncertain or consequential findings, enabling broad scenario exploration without relying on language models for numerical conclusions. A severe-weather workflow illustrates how these capabilities could identify vulnerabilities and evaluate mitigation options. The proposed CROFT provides the complementary research-to-operations pathway by evaluating the complete operational system, including operator interaction, data failures, cybersecurity, and integration with existing systems. Together, GridAI and CROFT support a progression from near-term advisory applications toward higher-consequence uses, with validation requirements increasing alongside operational risk.