Foundation Twins: A New Generation of Power Systems Digital Twins using Foundation AI Models

Technical Summary: Foundation Twins: A New Generation of Power Systems Digital Twins using Foundation AI Models

Problem Statement
Power systems are inherently multi-timescale systems, encompassing physical phenomena and decision-making processes that span from microseconds (e.g., frequency control, protections) to decades (e.g., expansion planning). Current Digital Twin (DT) concepts for power systems remain distant from practical implementation because they lack the capability to provide unified decision-making across these disparate timescales, horizons, and geographic scopes. Traditional approaches often rely on splitting system variables into separate sub-vectors (slow, fast, and intermediate dynamics) to manage complexity, resulting in fragmented models that fail to effectively transfer information, uncertainty, and decisions across timescales. Consequently, there is a critical gap between power system operation and planning frameworks, limiting system resilience. The paper argues that existing research has not yet delivered a DT capable of bridging this gap without relying on separate physics-based models that exploit dynamics separation.

Methodology: The Foundation Twins Architecture
The paper proposes a new concept called Foundation Twins, which leverages recent advances in Artificial Intelligence (AI) and Machine Learning (ML), specifically Foundation Models and Reinforcement Learning (RL). Rather than a single monolithic model, the proposed architecture consists of several specialized foundation AI models orchestrated by a central Manager, complemented by an RL framework for decision-making. The architecture includes the following core components:

1. Manager Module: Acts as the coordinator, receiving optimization tasks from operators and configuring the settings for other models. It interfaces with the physical infrastructure for measurements and control actions.
2. Time-Series Foundation Model: Responsible for generating time-series data at various required resolutions. It fills gaps in temporal or geographic data, provides forecasts for exogenous variables (e.g., renewable generation, demand), and ensures data follows physical phenomena while handling extreme values and conditioning constraints.
3. State Estimator Foundation Model: Estimates the current power system state at the appropriate timescale defined by the Manager. It utilizes time-series data to provide synchronous state estimations across different operational speeds.
4. Power System Foundation Model: Serves as a high-fidelity simulator of the physical power system. Its primary role is to predict future system states and rewards across multiple timescales and horizons, driven by proposed actions. It aims to represent multiple physics phenomena (transients, steady-state) and uncertainties within a single model.
5. Optimizer Module (RL Architecture): Implements an Actor-Critic structure. The Actor proposes actions based on the current state, while the Critic estimates future accumulated rewards based on state transitions predicted by the Power System Foundation Model. This module is designed to learn policies for multi-timescale decision-making, potentially utilizing Hierarchical RL to break long-term goals into sub-goals.
6. Memory Module: Stores state-action-reward transitions for future reference by the Actor and Critic.

Key Contributions
This paper is explicitly framed as a position paper rather than a standard research paper presenting experimental results. Its primary contributions are conceptual and visionary:

• Introduction of "Foundation Twins": A new paradigm for power system DTs that combines the generalization features of foundation models with the decision-making capabilities of RL.
• Architectural Proposal: A detailed blueprint (Fig. 2) outlining how multiple specialized foundation models (Time-Series, State Estimator, Power System) can be integrated with an RL optimizer to achieve multi-timescale decision-making in a single ecosystem.
• Identification of Critical Research Gaps: The paper systematically identifies the specific technical challenges required to realize this vision, including:
  • Developing unified generative architectures for time-series data that respect physics, handle multiple energy patterns, and model extreme values without model collapse.
  • Informing multi-timescale physics into ML models (e.g., via continuum representations or differentiable solvers) to ensure generalization and convergence.
  • Creating efficient state representations (e.g., hybrid sequence-graph models) that are invariant to power system configuration changes.
  • Solving Hierarchical RL challenges, such as designing feasible latent spaces for power system optimization and distributing long-term goals into sub-goals without losing explainability.

Results and Claims
As a position paper, the work does not present experimental results, simulations, or performance benchmarks. Instead, it claims that the proposed architecture addresses the limitations of current DTs by:

• Eliminating the need for separate physics-based models for different timescales.
• Enabling knowledge transfer (dynamics, uncertainty, decisions) across the entire spectrum of power system processes.
• Offering a path toward more resilient power systems by simultaneously accounting for fast and slow dynamics (e.g., coordinating DER dispatch with stability constraints).

The author notes that while some components (like the Manager and Memory modules) are considered less complex, the realization of the Foundation Twins depends heavily on overcoming significant scientific challenges in physics-informed ML, state representation, and hierarchical decision-making.

Significance
The paper positions "Foundation Twins" as a necessary evolution for power system digitalization. It argues that current DT concepts are stalled because they cannot handle the multi-timescale nature of power systems. By leveraging foundation models for simulation and generalization, combined with RL for decision-making, this approach aims to deliver the long-envisioned capability of a unified, high-fidelity digital twin. The paper serves as a roadmap for future research, urging the community to focus on the specific technical hurdles identified (such as physics-informed multi-timescale modeling and hierarchical RL coordination) to move power system DTs from a distant concept to a deployable reality. The author acknowledges that while the current proposal follows a centralized operator perspective, the concept could be extended to distributed decision-making approaches.