Foundation Twins: A New Generation of Power Systems Digital Twins using Foundation AI Models
Pedro P. Vergara · 2026
Read Foundation Twins: A New Generation of Power Systems Digital Twins using Foundation AI Models by Pedro P. Vergara (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-05952.txt.
Introduction
Pedro Vergara sets out a vision for digital twins in power systems, in a self described position paper. A digital twin is a model that mirrors a real grid and helps guide choices. Power grids act across many time scales and places, yet twin research has so far stayed far from real use, he argues.
Vergara proposes Foundation Twins, which join foundation AI models with reinforcement learning systems that support decisions. The foundation side brings broad generalization across cases, while the learning side handles choices at varied speeds and scopes. His aim is a twin that can speed and improve grid decisions rather than sit as a lab concept. The paper maps how such systems could be built.
Catalog record
| CREATOR | Pedro P. Vergara |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 30 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.05952
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Foundation Twins: A New Generation of Power Systems Digital Twins using Foundation AI Models Pedro P. Vergara Abstract Power systems are inherently multi-timescale systems, with different physical phenomena and decision-making processes spanning multiple timescales, time horizons, and geographic scopes. I envision power systems digital twins (DTs) as powerful modeling and simulation tools that can accelerate and improve decision-making across different time scales and geographic scopes. However, until now, research has not delivered such a vision, and power systems DTs remain a concept distant from implementation. This is not a regular research paper. This is a position paper that outlines my vision for developing a new generation of power systems DTs that leverage recent advances in artificial intelligence (AI) and machine learning (ML). I call these Foundation Twins . Foundation Twins combines the generalization features of foundation models with the decision-making capabilities of reinforcement learning (RL) architectures to deliver the envisioned power systems DTs. I Introduction Digital Twins (DTs), in the context of power systems, can be defined as a … ”collection of modules and models (based on multi-physics simulation) integrated into a single (software) ecosystem … aimed at mirroring the real-time operation of the power systems and supporting its long-term planning.” [ 17 ] . According to this definition, DTs are expected to be composed of several (multi-physics) simulation models that serve as the basis for other processes. Some such processes may include dispatching distributed energy resources (DERs), frequency and voltage control, and managing transformer and cable capacity. In general, these processes operate at different timescales, horizons, and geographic scopes, suggesting that DTs must provide multi-timescale and multi-geographic decision-making capabilities. Until now, such capabilities have been missing, and power systems DTs remain a concept distant from deployment. The need for multi-timescale decision-making is readily apparent from a practical perspective. For instance, a single DT that coordinates across multiple timescales could help dispatch DERs in the market (slow dynamics) while accounting for power system stability (fast dynamics), thereby avoiding, for example, voltage and frequency issues. As a result, by accounting simultaneously for fast and slow dynamics, DTs can bridge the current gap between power system operation and planning frameworks, thereby increasing its resilience (e.g., against blackouts). Fig. 1: Approximated power systems processes and their different timescales. Taking into account all power systems processes in a unique model would result in an untractable model. I-A Multi-Timescale Power Systems Power systems are composed of a set of devices (i.e., assets, resources) that can be described by continuous dynamics and discrete events. A simple example, synchronous machines, can be modeled using a set of non-linear ordinary differential equations (ODEs) with proper parameters. Other devices, such as controllers, can also be modeled using ODEs. In discrete events, Boolean or discrete variables can be used to model, for instance, switch connection statuses or tap values of tap-changing transformers. In power systems analysis, the time constants of different physical phenomena, as well as some decision-making processes, can span a wide range of different timescales and horizons. Fig. 1 presents the time scales of a variety of typical power system decision-making processes. Since the timescale spans from microseconds to decades, taking into account the dynamics of all phenomena and processes in a unique model would result in an untractable approach. A common solution to this issue is to split the general state vector of system variables s into three sub-vectors [ 11 ] : (i) state variables s_{I} characterized by slow dynamics, (ii) state variables s_{II} characterized by dynamics of interest, and (iii) state variables s_{III} characterized by fast dynamics. The consequence of this division is that, from the perspective of the state variables of interest s_{II} , the time evolution of state variables s_{I} can be considered so slow, while the time evolution of variables s_{III} can be considered so fast, that their variations can be disregarded. As a result, continuous dynamics can now be modeled using nonlinear differential-algebraic equations (DAEs) using both continuous and discrete variables. Although this modeling approach enables separate models for different physical phenomena and decision-making processes, the multi-timescale nature of power system infrastructure persists. Models that enable proper transfer of (fast and slow) dynamics, information, uncertainty, errors, and decisions across timescales are urgently needed. The question that remains is: How can we develop DTs that provide decision-making capabilities across multiple timescales, horizons, and geographic scopes, without relying on separate physics-based models that exploit dynamics separation? . I believe that recent developments in artificial intelligence (AI) and machine learning (ML), particularly in foundation models, complemented by the decision-making capabilities of reinforcement learning (RL), can help DTs deliver these expected features. Fig. 2: Foundation Twins architecture composed of several foundation AI models to enhance simulation capabilities, complemented by an RL architecture for decision-making. The Manager is responsible for coordinating all other models and modules, as well as for interfacing measurements and control actions with the power systems infrastructure. Measurements from the physical power system are fed into the Time-Series Foundation Model at different time resolutions. The Power Systems Foundation Model serves as a simulator, while the State Estimator Foundation Model uses time-series data to estimate the current system state. Based on the current state, the Optimizer Module proposes future actions using an Actor , while expected (accumulated) rewards are estimated via the Critic . Future system states are provided via the Power System Foundation Model . Finally, the Memory Model stores state-action transitions (and their rewards) for future reference by the Optimizer Module. I-B Foundation Models Foundation models have recently emerged as a new learning paradigm in AI. These models learn from large datasets through self-supervision and have proved to generalize across many applications. Unlike supervised learning models, foundation models can generalize across multiple learning tasks with a single architecture, often based on a transformer architecture with an encoder and decoder components. These components are important because they enable the foundation model to reconstruct masked parts of the input data, thereby allowing it to learn from unannotated data. Successful examples of foundation models include the now-commercial large language models (LLMs) such as ChatGPT and LeChat, but their applications extend to time-series data, audio, and video. The potential of foundation AI models for power systems has already been discussed in [ 4 ] , where features such as the ability to generalize across multiple tasks with a single architecture and to learn from multimodal data sources have been highlighted as crucial for power systems applications. I argue that although these aspects are important, crucial features required for power system applications are still missing (which I discuss here as well). Moreover, I also argue that a single (and large) model may not be sufficient to deploy the expected capabilities. Instead, I foresee the use of multiple specialized foundation models, properly orchestrated to enhance multi-timescale simulation capabilities, complemented by an RL architecture to deploy decision-making. This is the general idea behind the concept that I introduce in this paper. I-C Contributions This is not a regular research paper. This paper aims to outline a vision for developing a new generation of power systems DTs using recent advances in foundation AI models and RL. Many of the ideas described in this paper have been formulated by various authors