Predicting Power-System Dynamic Trajectories with Foundation Models
Haoran Li, Lihao Mai, Chenhan Xiao, Erik Blasch, and Yang Weng · 2026
Read Predicting Power-System Dynamic Trajectories with Foundation Models by Haoran Li, Lihao Mai, Chenhan Xiao, Erik Blasch, and Yang Weng (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-2604-14991.txt.
Introduction
Li and four others propose LASS-ODE-Power, a learning framework for general time-domain prediction of power-system dynamics. Existing learning approaches train per system and generalize poorly across conditions and parameters. Their model pretrains on over 40 GB of ODE trajectories, then predicts from short starts across electromechanical and inverter regimes, zero-shot without data sharing. Parallel linear computation is fast, and fine-tuning on about 1 GB of power-system data boosts task performance. Tests across diverse cases show better accuracy with efficient inference than existing models.
Catalog record
| CREATOR | Haoran Li, Lihao Mai, Chenhan Xiao, Erik Blasch, and Yang Weng |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53572 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.14991
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Predicting Power-System Dynamic Trajectories with Foundation Models Haoran Li, Lihao Mai, Chenhan Xiao, Erik Blasch, and Yang Weng Abstract As power systems transition toward renewable-rich and inverter-dominated operations, accurate time-domain dynamic analysis becomes increasingly critical. Such analysis supports key operational tasks, including transient stability assessment, dynamic security analysis, contingency screening, and post-fault trajectory evaluation. In practice, these tasks may operate under several challenges, including unknown and time-varying system parameters, privacy constraints on data sharing, and the need for fast online inference. Conventional high-fidelity prediction based on nonlinear differential–algebraic equations (DAEs) requires detailed domain knowledge and can be computationally expensive. Meanwhile, existing learning-based approaches are typically trained for individual systems and therefore lack generalization across operating conditions and physical parameters. Hence, this paper proposes LArge Scale Small ODE (LASS)-ODE-Power, a learning framework for general-purpose time-domain prediction. The proposed approach leverages large-scale pretraining on more than 40 GB of DAE or ordinary differential-equation (ODE) trajectories to learn transferable representations. The resulting model supports trajectory prediction from short measurement prefixes across diverse dynamic regimes, including electromechanical and inverter-driven systems. Hence, the model can be directly used without data sharing in a zero-shot setting. In addition, the proposed architecture incorporates parallel and linearized computation to achieve fast inference. Moreover, to enhance task-specific performance in power systems, a specialized fine-tuning strategy is developed based on approximately 1 GB of heterogeneous power-system dynamic data. Extensive experiments over diverse power-system simulation scenarios demonstrate that LASS-ODE-Power consistently outperforms existing learning-based models in trajectory prediction accuracy with efficient inference. I Introduction The electric power grid is undergoing a structural transition driven by high penetrations of inverter-based resources (IBRs) [ 2 ] , expanding AC/DC interconnections [ 41 ] , and increasingly variable operating conditions [ 20 ] . These changes reduce operational margins and increase the need for accurate time-domain dynamic analysis. Such analysis supports several critical tasks in power system operation, including transient stability assessment [ 19 , 46 ] , dynamic security assessment (DSA) [ 44 ] , and dynamic state estimation [ 35 , 37 , 23 ] . It is also important for contingency screening [ 51 ] and post-fault trajectory analysis [ 25 , 26 , 24 ] . In modern grid operations, such simulations must be executed repeatedly across many operating points, topology configurations, and disturbance scenarios under strict time constraints [ 44 , 19 , 27 ] . As a result, scalable and efficient time-domain simulation and prediction have become increasingly important for both planning and real-time operational studies. In practice, this goal introduces several challenges. First, dynamic-model parameters are often uncertain or time-varying due to changing operating conditions, controller updates, etc. In addition, fault events, protection actions, and switching operations induce discrete mode transitions that produce piecewise hybrid trajectories that are difficult to capture using a single fixed-parameter model [ 49 , 21 ] . Second, measurement data-quality issues are pervasive, including noise, missing channels, asynchronous sampling, and partial observability, which can degrade numerical stability and the generalization performance of learned surrogates [ 52 , 29 , 36 ] . Third, privacy and proprietary constraints often prevent operators and vendors from sharing measurements and device parameters in centralized training environments [ 9 ] . Finally, practical studies require evaluating large ensembles of disturbance scenarios and operating conditions within tight latency budgets, making high-throughput and fast inference essential [ 8 , 6 ] . Mathematically, time-domain dynamic prediction corresponds to solving nonlinear initial-value problems described by differential-algebraic equations (DAEs) that couple device-level differential dynamics with algebraic network constraints. In renewable-rich and inverter-dominated grids, the resulting dynamics exhibit pronounced multi-timescale behavior and hybrid switching events. Practical studies may therefore rely on electromagnetic-transient (EMT) simulation or EMT-transient-stability (TS) hybrid/co-simulation with explicit event handling [ 47 , 31 ] . Increasingly, such simulations are aligned with field measurements to capture real-time system dynamics and calibrate models using heterogeneous monitoring streams with different sampling rates and resolutions [ 38 , 22 ] . Classic high-fidelity numerical simulation and prediction remain the gold standard for analyzing power-system dynamics when accurate parameters and models are available. For example, EMT simulation and EMT-TS co-simulation provide highly detailed representations of dynamic behavior [ 18 ] . However, these approaches are computationally expensive because they repeatedly solve nonlinear network equations and perform fine-step numerical integration. Consequently, evaluating large ensembles of disturbance scenarios can become throughput-limited even when accelerated using GPU or high-performance computing techniques [ 50 ] . Moreover, uncertain parameters, repeated initialization procedures, and time-varying operating conditions can further degrade simulation fidelity. To reduce computational costs and improve modeling under uncertain parameters, learning-based approaches for dynamic prediction have recently attracted attention. Symbolic regression methods attempt to recover governing equations directly from trajectory data, for example, using genetic programming [ 45 ] or sparse identification techniques such as SINDy [ 4 ] . While these approaches can produce interpretable models when the search space is well designed and data quality is high, they are often sensitive to noise, missing measurements, and derivative estimation errors, and they scale poorly to large and high-dimensional systems. Another major family of methods is physics-informed learning, such as physics-informed neural networks (PINNs), which enforce DAE residual constraints during training to promote physically consistent prediction and parameter estimation [ 43 , 17 ] . Several recent studies have explored such techniques in power-system applications [ 23 , 37 , 28 ] . However, enforcing physical consistency typically requires partial knowledge of the underlying DAE structure, which can limit portability across heterogeneous devices and modeling conventions. Closely related continuous-time approaches, such as neural differential equation models [ 7 , 42 ] , parameterize system dynamics using neural networks and learn them through differentiable ODE solvers. Although these models can approximate complex nonlinear dynamics without requiring explicit parameters, training and inference remain computationally expensive due to repeated numerical integration. In addition, such models are often trained for individual systems, making cross-system generalization difficult. These limitations motivate the development of reusable learning models that can generalize across systems, remain robust under realistic measurement imperfections, preserve privacy constraints, and enable high-throughput time-domain prediction without costly nonlinear integration. Achieving these goals simultaneously is challenging, particularly when models are trained on system-specific datasets with limited coverage of dynamic behaviors. A promising direction is to learn transferable representations of dynamic trajectories from large historical datasets. Inspired by recent advances in large-scale pretraining across machine learning domains, our prior work proposed LA rge S cale S mall ODE (LASS-ODE) [ 30 ] , a reusable model for learning dynamic systems from diverse ODE/DAE trajectories. LASS-ODE supports two key capabilities: online parameter adaptation and trajectory prediction conditioned on short observed measurement segments. Through large-scale pretraining on more than 40 GB of DAE trajectories, the model learns general dynamic patterns and improves robustness under measurement noise and missing data. In addition, the architecture incorporates linearized and parallel computation to enable efficient training and inference. In this work, we instantiate this paradigm for power systems and develop LASS-ODE-Power , a foundation model for measurement-based adaptive time-domain prediction in power grids. The pretrained LASS-ODE model is further fine-tuned on approximately 1 GB of multi-timescale power-system dynamic data to adapt it to grid-specific dynamics. A widely used adaptation strategy is