Trusted GridAI

Session Chair: Spyros Chatzivasileiadis – DTU

This session focuses on:

• the industry needs and
• the methods and tools to make AI trustworthy and safe to be deployed in power systems.

Speakers will address the needs for benchmarking, validation and verification of AI methods, and will offer directions and tools that can address these needs.

Grid Foundation Models: Applications and Requirements

Speaker: Jan Poland – Hitachi Energy Research

This contribution provides inputs on how Grid Foundation Models might start providing real value in the near future, from an industry perspective. We will highlight some (downstream) applications that we expect to be good first use cases, the main enabling properties which we expect will facilitate adoption, some key requirements on Foundation Models we see in order to be accepted by industry, and several challenges that still need to be addressed. We will discuss (synthetic) training data, trust, validation, performance criteria, and other aspects.

A Trustworthiness Layer for GridFM via Conformal Prediction Intervals: Implementation and Open Questions

Speaker: Antonio Alcántara Mata – CUNEF Universidad

Foundation models such as GridFM predict AC grid states orders of magnitude faster than iterative solvers, making systematic N-k contingency screening and extensive power system analysis computationally feasible. Speed alone is not enough for operational use: a plausible but wrong “”safe”” prediction during a genuine overload is precisely the failure operators cannot tolerate. GridFM is missing a way to attach quantified, statistically valid uncertainty to its outputs.

We present gridfm-trustlayer, an open-source, model-agnostic layer that equips GridFM predictions with distribution-free prediction intervals via conformal prediction. It operates post-hoc on prediction artifacts, so it requires no retraining and no access to model internals, and its statistical core is deliberately independent of the training stack. The layer offers two calibration families: stratified conformal prediction, which groups calibration residuals by contingency severity and other scenario descriptors; and kernel-weighted conformal prediction, which localizes calibration to each test scenario using the model’s own latent representations to obtain approximately conditional bounds. Both can feed a screening stage that turns calibrated intervals into operational violation alarms, or be used as a proxy for model uncertainty.

The talk covers the current implementation, design decisions, and open questions for the trustworthiness layer.”

Trusted Grid AI

Speaker: Iason Langford – Energinet

The speaker will address the needs for benchmarking, validation and verification of AI methods, and will offer directions and tools that can address these needs.

Future Industry needs for AI Trustworthiness

Speaker: Venkat Banunarayanan – National Rural Electric Cooperative Association (NRECA)

As AI initiatives move from pilot projects into operational deployment, trustworthiness is a critical operational requirement. Organizations need to show all stakeholders such as regulators, consumers, and their own staff that the AI behaves reliably, transparently, and safely under real-world conditions. This presentation will illustrate key aspects of trustworthiness such as continuous assurance in place of one-time certification, evaluation methods that keep pace with rapidly evolving foundation models, accountability needs for AI agents, transparency across AI models, and the need to explore harmonization across fragmented global regulations. Trust is becoming both a requirement and a competitive differentiator, and meeting future needs will require closer collaboration between developers, deployers, and policymakers.