Poster 1: SABLE_PA: Fully GPU-Accelerated Learnable Sparse Batched Power Flow Accelerator

Authors: Suho Park; Keunju Song; Hongseok Kim*
Contact: suho Park qkrtngh8210@gmail.com
Organization: NICELAB, Department of Electrical Engineering, Sogang University

Abstract:
Modern power systems require real-time analysis of massive grid states. We propose SABLE PA (Sparse Accelerated Batched Learnable Power Flow Analysis), a fully GPU-accelerated solver designed for both numerical efficiency and deep learning integration.First, a Virtual Block-Diagonal Embedding (VBDE) architecture maps batched data into a unified 2D sparse structure, enabling seamless integration as a differentiable layer within Physics-Informed Neural Networks (PINNs). Second, a mixed-precision scheme using NVIDIA cuDSS—utilizing FP32 for linear solves while maintaining FP64 accuracy for residuals—significantly reduces expensive memory operations. Third, sparsity-preserving in-place update and block-aware indexing kernels maximize throughput by eliminating redundant indexing overhead.Benchmarks evaluate SABLE PA as both a standalone solver and an integrated implicit layer. For a 6495rte system (batch=128), SABLE PA achieves a 783x times speedup over pandapower (CPU) and a 4x times speedup over ExaPF (2025) (GPU). When integrated into the DeepLDE framework for a 2312-bus system, it accelerates overall training by 36.5x times and the implicit solving process by 48.4x times. These results demonstrate SABLE PA’s efficacy as a rapid, memory-efficient layer for restoring physical feasibility in constrained learning pipelines.

Poster 2: Type-Aware Latent Graph Neural ODEs for Sensor-Dependent Imputation in Power Distribution

Authors: Pooja Algikar, Stefano Fenu, Kumar Jhala, Siby Jose Plathottam
Contact: Pooja Algikar palgikar@anl.gov
Organization: Argonne National Lab

Abstract:
Modern distribution systems are increasingly instrumented by heterogeneous sensors (e.g., PMUs, SCADA, DER monitors, smart meters) that report different subsets of electrical variables at different sampling rates and with structured missingness. While “heterogeneous data fusion” is often framed as multi-modal learning, in power distribution the heterogeneity is frequently measurement-operator–induced: each node is associated with a sensor type that determines which state components are observable and how they are mapped into a common system representation. We propose a type-aware latent graph neural ODE for distribution-system data fusion and imputation, where heterogeneity is modeled explicitly at the level of (i) node embeddings, (ii) message passing, and (iii) reconstruction. Our encoder learns node representations using sensor-type–conditioned feature slicing and adaptation layers, enabling shared latent geometry while respecting sensor-specific observability. To model continuous-time evolution under irregular sampling, we parameterize the latent dynamics with a graph neural vector field, optionally augmented with typed attention to modulate information exchange across sensor types. Finally, a type-specific decoder implements the inverse measurement mapping by reconstructing only the appropriate feature subspaces per node type and scattering outputs into a global catalog, yielding faithful reconstructions even under partial observability. Experiments on distribution-system power-flow data with controlled observation ratios demonstrate improved imputation accuracy and robustness to missingness patterns compared to homogeneous latent ODE baselines, while providing a principled mechanism to incorporate sensor-dependent measurement structure into continuous-time graph generative modeling.

Poster 3: Operator Learning for Power Systems Simulation

Authors: Matthew Schlegel, Matthew E. Taylor, Mostafa Farrokhabadi
Contact: Matthew Schlegel matthew.schlegel@ucalgary.ca
Organization: University of Calgary, Schulich School of Engineering

Abstract:
Time domain simulation, i.e., modeling the system’s evolution over time, is a crucial tool for studying and enhancing power system stability and dynamic performance. However, these simulations become computationally intractable for renewable-penetrated grids, due to the small simulation time step required to capture renewable energy resources’ ultra-fast dynamic phenomena in the range of 1-50 microseconds. This creates a critical need for solutions that are both fast and scalable, posing a major barrier for the stable integration of renewable energy resources and thus climate change mitigation. This paper explores operator learning, a family of machine learning methods that learn mappings between functions, as a surrogate model for these costly simulations. The paper investigates, for the first time, the fundamental concept of simulation time step-invariance, which enables models trained on coarse time steps to generalize to fine-resolution dynamics. Three operator learning methods are benchmarked on a simple test system that, while not incorporating practical complexities of renewable-penetrated grids, serves as a first proof-of-concept to demonstrate the viability of time step-invariance. Models are evaluated on (i) zero-shot super-resolution, where training is performed on a coarse simulation time step and inference is performed at super-resolution, and (ii) generalization between stable and unstable dynamic regimes. This work addresses a key challenge in the integration of renewable energy for the mitigation of climate change by benchmarking operator learning methods to model physical systems.

Poster 4: Universal Graph Learning Algorithms for Power Network Systems: Zero-Shot Transferability and Beyond

Authors: Tong Wu; Anna Scaglione; Sandy Miguel; Daniel Arnold
Contact: Tong Wu tong.wu@ucf.edu
Organization: University of Central Florida

Abstract:
This work addresses a fundamental challenge in applying deep learning to power systems: developing neural network models that transfer across significant system changes, including networks with entirely different topologies and dimensionalities, without requiring training data from unseen reconfigurations. Despite extensive research, most ML-based approaches remain system-specific, limiting real-world deployment. This limitation stems from a dual barrier. First, topology changes shift feature distributions and alter input dimensions due to power flow physics. Second, reconfigurations redefine output semantics and dimensionality, requiring models to handle configuration-specific outputs while maintaining transferable feature extraction. To overcome this challenge, we introduce a Universal Graph Convolutional Network (UGCN) that achieves transferability to any reconfiguration or variation of existing power systems without any prior knowledge of new grid topologies or retraining during implementation. Our approach applies to both transmission and distribution networks and demonstrates generalization capability to completely unseen system reconfigurations, such as network restructuring and major grid expansions. Experimental results across power system applications, including false data injection detection and state forecasting, show that UGCN significantly outperforms state-of-the-art methods in cross-system zero-shot transferability of new reconfigurations.

Poster 5: When AI Meets the Grid: Reliable and Sustainable “AI-Power Nexus”

Authors: Xin Chen
Contact: Xin Chen xin_chen@tamu.edu
Organization: Texas A&M University

Abstract:
AI is expanding at an extraordinary pace and placing unprecedented demands on electric power systems. To support reliable grid operations and accelerated AI development, my research advances the “AI-Power Nexus” along two closely connected directions. First, “AI for Power (A4P)” develops large language model powered agentic AI that interacts with power system databases and professional tools to assist with complex grid operations and decision making. These AI agents autonomously formulate grid studies, run power flow and stability analyses, explore contingencies, interpret results, and generate clear documentation, while keeping human operators firmly in the loop for oversight and final decisions. Second, “Power for AI (P4A)” focuses on high-fidelity dynamic modeling and advanced control of AI data centers. We develop modularized dynamic models of data center electrical architectures, study their behavior across time scales, and design grid-friendly control strategies that enhance voltage support, disturbance ride-through capability, and effective coordination with the power grid. By interconnecting “A4P” and “P4A”, this research outlines a path toward AI-ready and AI-enabled grids that are more reliable, efficient, and resilient.

Poster 6: Graph Neural Networks for Large-Scale Power Flow Calculations in Electric Grids

Authors: Mohammed Olama (ORNL), Massimiliano Lupo Pasini (ORNL), Ajay Yadav (ORNL), Ali Trigui (Qubit Engineering), and Marouane Salhi (Qubit Engineering)
Contact: Mohammed Olama olamahussemm@ornl.gov
Organization: Oak Ridge National Laboratory

Abstract:
Accurate and efficient power flow calculations are critical for modern grid management. Traditional numerical solvers face scalability challenges, especially in large and dynamic power networks. To address this challenge, Oak Ridge National Laboratory (ORNL), in collaboration with Qubit Engineering and the Tennessee Valley Authority (TVA), have developed a cutting-edge physics-informed artificial intelligence (AI)-based heterogeneous Graph Neural Network (GNN) solution tailored for AC Optimal Power Flow, optimized specifically for TVA’s grid topology and trained using measured and synthetic data on ORNL’s high-performance computing (HPC) facilities. Our model learns complex system dynamics, enabling rapid and reliable predictions with reduced computational costs. Results demonstrate significant speedup while maintaining high accuracy, making GNNs a promising tool for real-time grid optimization and control.

Poster 7: Mid-Term Load Forecasting with Minimal Data: An In-Context-Learning-Aware Approach Using Large Language Models

Authors: Yating Zhou(Cornell University); Shuai Zhang(New Jersey Institute of Technology); Meng Wang(Rensselaer Polytechnic Institute)
Contact: Yating Zhou yz3554@cornell.edu
Organization: Cornell University

Abstract:
Accurate mid-term load forecasting at the building level is vital for the strategic planning, operation, and sustainability of modern power systems. Machine learning approaches often require large amounts of historical data for training, which may be unavailable in practice. Existing solutions typically rely on transfer learning, which requires extensive data from source domains and task-specific fine-tuning. These approaches are unsuited for mid-term forecasting, where long-range historical data is often limited. To address these challenges, this paper proposes Load-Context, a novel mid-term load forecasting method that leverages the in-context learning (ICL) capability of pre-trained large language models (LLMs). Load-Context consists of a frozen LLM backbone and a lightweight, trainable ICL-aware adapter, enabling effective few-shot and zero-shot forecasting without the need for fine-tuning. We further design a prompt construction strategy that captures spatial correlations among buildings and periodic load patterns. Once pre-trained, the proposed method can generalize to a wide range of target tasks through prompt-based adaptation. Experiments on real-world datasets demonstrate that Load-Context achieves high forecasting accuracy with minimal data.

Poster 8: Day-Ahead and Intraday Coordinated IES P2P Trading Strategy for Multi-Type Urban Entities with Bidding Consensus

Authors: Haotian Yang; Xinyue Chang; Yixun Xue; Xingtao Tian; Jiahe Xu; Changfu You; Junfu Lyu; Hongbin Sun
Contact: Haotian Yang haotian.yang.th@dartmouth.edu
Organization: thayer school of engineering

Abstract:
To address the impact of renewable energy output fluctuations on the energy distribution of a city‘s integrated energy system, this paper proposes a method that combines Nash bargaining and double-sided auctions to manage energy allocation in both the day-ahead and intraday stages. In the day-ahead stage, the Nash bargaining method designs exclusive trading strategies for each transaction, determining the trading periods, prices, and transaction volumes for all users. The formulated energy consumption plan can achieve optimal utility for all users while ensuring that each user receives a fair share of the benefits. In the intraday stage, when the day-ahead trading plan cannot be fulfilled due to energy output shortages, a new trading plan is needed, which triggers the initiation of the auction. The auction mechanism addresses the lengthy computation time of Nash bargaining. This makes the auction particularly suitable for intraday trading. The results demonstrate that the auction provides a favorable trade-off, offering energy management capability that is acceptably close to Nash bargaining while being dramatically faster. Moreover, the day-ahead trading price serves as a consensus, refining auction bidding strategies and facilitating the success of the auction. We also provide the criteria for enabling the auction, along with an explanation of the rationale behind the activation standers

Poster 9: Risk-Aware Market Clearing Mechanism for Data Center Capacity Allocation in Transmission Networks

Authors: Shaoze Li; Cong Chen
Contact: Shaoze Li shaoze.li.th@dartmouth.edu
Organization: Dartmouth College

Abstract:
Artificial-intelligence-driven data-center expansion is creating unprecedented electricity demand, while the available hosting capacity of existing power networks remains limited and grid reinforcement cannot be completed in the short term. We study how to characterize the guaranteed hosting capacity of the current network structure and how additional withdrawal capacity can be safely unlocked under a specified probability of risk. To achieve this, we propose a hierarchical robust model and a probabilistic risk-aware hosting model that maximize both total system capacity and nodal demand satisfaction. This approach demonstrates that tolerating a minimal, mathematically quantified blackout risk can release substantial hosting capacity under the existing grid infrastructure. Furthermore, to effectively distribute these scarce energy resources, we design a Simultaneous Ascending Auction (SAA) mechanism. We present the theoretical properties of the proposed SAA, specifically proving its convergence to a competitive equilibrium when bidders possess additive, unit-demand, $k$-demand, or symmetric concave valuation structures. Finally, numerical experiments validate the practical feasibility and effectiveness of our proposed allocation framework.

Poster 10: Urban infrastructure and fossil-fuel industrial legacy drive U.S. data center siting

Authors: Camilla Ancona, Center for Urban Science and Progress, New York University Tandon School of Engineering, 370 Jay St, Brooklyn, 11201, NY, USA; Ofek Lauber, Center for Urban Science and Progress, New York University Tandon School of Engineering, 370 Jay St, Brooklyn, 11201, NY, USA; Anton Rozhkov, Center for Urban Science and Progress, New York University Tandon School of Engineering, 370 Jay St, Brooklyn, 11201, NY, USA; Maurizio Porfiri, 1Center for Urban Science and Progress, New York University Tandon School of Engineering, 370 Jay St, Brooklyn, 11201, NY, USA Department of Mechanical and Aerospace Engineering, Department of Biomedical Engineering, and Department of Civil and Urban Engineering, New York University Tandon School of Engineering, 6 Metro Tech Center, Brooklyn, 11201, NY, USA.
Contact: Camilla Ancona ca3429@nyu.edu
Organization: Center for Urban Science and Progress, NYU Tandon School of Engineering

Abstract:
Data centers form the physical backbone of the digital economy, yet their integration into the United States’ urban fabric remains poorly understood. Challenging prevailing narratives of “rural cloud expansion,” we demonstrate that data centers are an overwhelmingly urban phenomenon, with 97.5% of facilities situated in cities. Their spatial logic is driven by the nameplate capacity of local power plants and, to a lesser extent, by broadband quality, retired coal plants, IT employment, and natural hazards. The positive association with the presence of retired coal plants highlights the role of fossil-fuel industrial legacy and existing transmission infrastructure in shaping contemporary digital development. Moreover, facilities under development are disproportionately located in Coal Closure Energy Communities, suggesting that areas affected by fossil-fuel plant retirements are emerging as strategic hosts for new digital infrastructure. Reliance on local power systems embeds data center siting within state-level energy generation patterns, amplifying geographic contrasts in the sustainability of the data center boom. Population-based projections of data centers’ growth indicate that the digital transition will reinforce existing urban hierarchies rather than catalyze decentralization. These findings suggest a path-dependent evolution linked to existing infrastructure and fossil-fuel industrial legacy, posing significant challenges to urban livability.

Poster 11: PowerAgent: A Road Map Toward Agentic Intelligence in Power Systems

Authors: Qian Zhang, Karina Chung, Skyler Liu, Le Xie
Contact: Qian Zhang qianzhang@g.harvard.edu
Organization: Harvard University

Abstract:
PowerAgent is the first open-source community dedicated to accelerating the development of Agentic AI in the power systems domain, which aim to bridge the gap between cutting-edge AI and the real-world needs of system operators, electric utilities, and researchers. The conceptual architecture, implementation pathways, and system-level benefits of deploying PowerAgent in power grid are discussed. Some ongoing application of deploying PowerAgent to speed up transmission planning process is also shared in the poster.

Poster 12: Mean-Field Learning for Storage Aggregation

Authors: Jingguan Liu (Huazhong University of Science and Technology); Cong Chen (Dartmouth College); Xiaomeng Ai (Huazhong University of Science and Technology); Jiakun Fang (Huazhong University of Science and Technology); Jinsong Wang (HyperStrong Technology Co., LTD); Jinyu Wen (Huazhong University of Science and Technology)
Contact: Jingguan Liu jingguan.liu.th@dartmouth.edu
Organization: Huazhong University of Science and Technology; Visiting Student at Dartmouth College

Abstract:
Distributed energy storage devices can be pooled and coordinated by aggregators to participate in power system operations and market clearings. This requires representing a massive device population as a single, tractable surrogate that is computationally efficient, accurate, and compatible with market participation requirements. However, surrogate identification is challenging due to heterogeneity, nonconvexity, and high dimensionality of storage devices. To address these challenges, this paper develops a mean-field learning framework for storage aggregation. We interpret aggregation as the average behavior of a large storage population and show that, as the population grows, aggregate performance converges to a unique, convex mean-field limit, enabling tractable population-level modeling. This convexity further yields a price-responsive characterization of aggregate storage behavior and allows us to bound the mean-field approximation error. Leveraging these results, we construct a convex surrogate model that approximates the aggregate behavior of large storage populations and can be embedded directly into power system operations and market clearing. Surrogate parameter identification is formulated as an optimization problem using historical market price-response data, and we adopt a gradient-based algorithm for efficient learning procedure. Case studies validate the theoretical findings and demonstrate the effectiveness of the proposed framework in approximation accuracy, data efficiency, and profit outcomes.

Poster 13: Communication-Limited Multi-Agent Grid Congestion Management via Differentiable Optimization

Authors: James Chen (MIT); Stéphane Drobot (RTE); Lucas Saludjian (RTE); Patrick Panciatici (RTE); Pin-Yu Chen (IBM Research); Ali Jadbabaie (MIT); Priya Donti (MIT)
Contact: James Chen jamesyc@mit.edu
Organization: MIT EECS/LIDS

Abstract:
While there have been many recent advances in multi-agent grid control, they often rely on more communication infrastructure than is available in existing real-world transmission systems. To address this, we present a novel framework for multi-agent grid congestion management inspired by the communication architecture used by RTE, the French transmission system operator. Our framework considers communication-limited architectures in which a central coordinator is only able to periodically provide information to an ensemble of local optimization-based controllers; these local controllers do not communicate, but are nonetheless dynamically coupled. The goal of the coordinator is to provide signals that limit adverse interactions between local controllers, and to ensure that the system as a whole minimizes operational costs and violations of thermal limits. To do so, we leverage the structure of the congestion management problem to expose a parameterized family of locally-verifiable constraints for each controller that jointly implies global constraint satisfaction. The central coordinator then must assign constraint values from this family to the local controllers, with the goal of minimizing the resultant cost over the joint closed-loop trajectory under forecasted scenarios of future disturbances. We frame this as a bi-level problem, and use recent advances in differentiable optimization to find an approximate solution. We demonstrate our method on an IEEE 118-bus system partitioned into three control areas. Our results demonstrate that our approach significantly lowers cost compared to RTE’s existing baseline. Furthermore, despite the significant communication constraints, our framework achieves costs comparable to the optimal cost under perfect information of future disturbances and with no restrictions on communication.

Poster 14: Certified Dispatchable Regions for Unbalanced Three-Phase Distribution Networks via Exact SDP Relaxation and Bundle Methods

Authors: Bohang Fang; Changhong Zhao (CUHK); Yue Chen (CUHK); Cong Chen (Dartmouth); Lijun Ding (UCSD)
Contact: Bohang Fang Bohang.Fang@dartmouth.edu
Organization: Dartmouth College

Abstract:
Dispatchable regions (also referred to as dynamic operating envelopes, DOEs) quantify the range of DER injections that a distribution network can safely accommodate while satisfying power flow feasibility and operational constraints. Constructing such regions for unbalanced three-phase networks remains challenging due to strong phase coupling and the nonlinear, nonconvex nature of AC power flow, especially when voltage and line safety limits must be enforced simultaneously. This work formulates dispatchable region characterization using a semidefinite programming (SDP) relaxation of the three-phase AC model. To address the long-standing concern of relaxation inexactness, we propose a penalty-based formulation that augments the objective with network power loss (together with slack variables for constraint handling), and we provide theoretical conditions under which the convex relaxation is guaranteed to be exact and yields physically meaningful, rank-one solutions. This establishes a principled pathway to obtain certified feasible operating envelopes for unbalanced feeders. To enable scalability toward feeder-scale, large-network settings where generic SDP solvers can become computationally prohibitive, we further explore a bundle method based solver framework that leverages cutting-plane models and fast eigen-computation to accelerate the SDP workflow. Ongoing work focuses on adaptive parameter tuning and structure-exploiting implementations to deliver operator-ready DOE computation with both rigor and speed.

Poster 15: Assessing the Impact of Electricity Markets on Resilience in the Context of Capacity Expansion: An Integrated Simulation-Optimization Approach

Authors: Weijie Pan (Dartmouth College); Ekundayo Shittu (George Washington University)
Contact: Weijie Pan weijie.pan@dartmouth.edu
Organization: Dartmouth College

Abstract:
Enhancing power system resilience to the growing frequency of high-impact, low-probability (HILP) contingencies is becoming increasingly urgent. As the integration of renewable energy (RE) accelerates global decarbonization efforts, challenges such as resource intermittency, substantial capital investments, and evolving market dynamics complicate the planning and construction of resilient power systems. Despite technological advancements, the lack of a holistic resilience assessment that considers both technical and non-technical dimensions hinders progress toward more integrated energy systems. This study evaluates the impact of electricity market structures on post-disaster system operations, aiming to identify optimal market mechanisms under expanded RE generation capacities. A sequential simulation-optimization approach is used, combining a Modelica-based tool to simulate the frequency of the system with the optimization of post-disaster operations in nine market modes. A multidimensional resilience assessment framework is proposed. Key findings from the modeling results include: (1) Expanded RE generation capacity can destabilize system frequency, but it mitigates energy supply risks during post-disaster periods; (2) Electricity market modes significantly influence the unit cost of purchasing solar energy from prosumers, with real-time markets initially reducing costs, but day-ahead markets becoming more effective as RE capacity increases; and (3) A threshold RE capacity proportion of 13.64% is identified, where resilience performance reaches its peak, shifting the preferable market mode from real-time to day-ahead beyond this point. This study contributes methodologically by proposing an integrated modeling approach and assessment metrics for evaluating system resilience that consider both technical performance and economic costs. Practically, it provides insights for system planners and policymakers on balancing RE expansion with market dynamics to enhance system resilience.

Poster 16: Online Smoothed Demand Management

Authors: Adam Lechowicz, University of Massachusetts Amherst;
Nicolas Christianson, Stanford University;
Mohammad Hajiesmaili, University of Massachusetts Amherst;
Adam Wierman, California Institute of Technology;
Prashant Shenoy, University of Massachusetts Amherst
Contact: Adam Lechowicz alechowicz@umass.edu
Organization: University of Massachusetts Amherst

Abstract:
We introduce and study online smoothed demand management (OSDM), motivated by paradigm shifts in grid integration and energy storage for large energy consumers (e.g., data centers). In OSDM, an operator makes two decisions at each time step: an amount of energy to be purchased, and an amount of energy to be delivered (i.e., used for computation). The difference between these decisions charges/discharges the operator’s energy storage. Two types of demand arrive online: base demand, which must be covered at the current time, and flexible demand, which can be satisfied at any time before a demand-specific deadline. The operator’s goal is to minimize a cost (subject to above constraints) that combines a cost of purchasing energy, a cost for delivering energy (if applicable), and smoothness penalties on the purchasing and delivery rates to encourage “grid healthy” decisions. We propose an algorithm for OSDM and show it achieves the optimal competitive ratio. To overcome the pessimism typical of worst-case (competitive) analysis, we also propose a novel framework that allows end-to-end differentiable learning of the best algorithm on historical problem instances while maintaining worst-case guarantees (i.e., to provide robustness against nonstationarity). We evaluate our algorithms in a case study of a grid-integrated data center with battery storage, showing that our algorithms effectively solve the problem and end-to-end learning achieves further substantial performance improvements.

Poster 17: LUMINA: A Grid Foundation Model for Benchmarking AC Optimal Power Flow Surrogate Learning

Authors: Hongwei Jin (Argonne National Laboratory); Keunju Song (Sogang University); Zeeshan Memon (Emory University); Yijiang Li (Argonne National Laboratory); Stefano Fenu (Argonne National Laboratory); Hongseok Kim (Sogang University); Liang Zhao (Emory University); Kibaek Kim (Argonne National Laboratory)
Contact: Yijiang Li yijiang.li@anl.gov
Organization: Argonne National Laboratory

Abstract:
AC optimal power flow (ACOPF) is foundational yet computationally expensive in power grid operations, driving learning-based surrogates for large-scale grid analysis. These surrogates, however, often fail to generalize across network topologies, a critical gap for deployment on grids not seen during training and for routine operational what-if studies. We introduce LUMINA-Bench, a comprehensive benchmark suite for ACOPF surrogate learning covering multi-topology pretraining, transfer, and adaptation. The benchmark evaluates homogeneous and heterogeneous architectures under single- and multi-topology learning settings using unified metrics that capture both predictive accuracy and physics-informed constraint violations. We additionally compare constraint-aware training objectives, including MSE, augmented Lagrangian, and violation-based Lagrangian losses, to characterize accuracy–robustness trade-offs across settings. Data processing, training, and evaluation frameworks are open-sourced as the LUMINA suite to support reproducibility and accelerate future research on feasibility-aware OPF surrogates.

Poster 18: Zero-Shot Power System Anomaly Detection and Mitigation Using Large Language Models

Authors: Saleh Sadeghi; Wei Sun
Contact: Saleh Sadeghi saleh.sadeghi@ucf.edu
Organization: University of Central Florida

Abstract:
Modern electric grids are evolving into large, data-intensive cyber-physical systems that generate massive volumes of multivariate time series measurements across geographically distributed networks. Detecting cyber and physical anomalies in such high-dimensional and dynamic environments remains a critical challenge, particularly because labeled data for rare faults and attacks are limited. Traditional machine learning approaches often require retraining or task-specific supervision, which restricts scalability and adaptability. These challenges motivate the exploration of foundation models that can generalize across tasks and operate without domain specific fine tuning. This study presents ZS-LAD, a zero-shot large language model framework for anomaly detection and mitigation in power systems. A frozen large language model is used as a general sequence predictor to forecast grid measurements without any training on power system data. Forecast residuals are evaluated using robust statistical analysis to detect abnormal behavior. Detected events are then interpreted by the same model to generate structured event classification and operator-level mitigation recommendations. Evaluation of the IEEE 13-bus feeder using a hardware-in-the-loop co-simulation platform with an OPAL-RT real-time simulator demonstrates reliable detection, low response time, and clear interpretability. The results highlight how foundation models can support scalable, adaptable, and reasoning driven grid monitoring in modern electric networks.

Poster 19: Randomness as a Resource for Electric Power Systems

Authors: Samuel Talkington; Dmitrii M. Ostrovskii; Daniel K. Molzahn
Contact: Samuel Talkington talkington@gatech.edu
Organization: Georgia Tech

Abstract:
What if we could better manage society’s electric power systems by transforming randomness from uncertainty to be managed into a resource to be leveraged? In this poster, I show how tools from randomized algorithms can be unified with power flow physics to create efficient and reliable approximation algorithms for network reconfiguration problems. Central to this framework is the insight that electrical congestion can be captured via graph Laplacian quadratic forms, enabling matrix concentration inequalities to provide guarantees for randomized switching decisions. This approach achieves orders-of-magnitude speedups over state-of-the-art commercial solvers, with comparable solution quality.

Poster 20: Forecasting AC-OPF with GridFM

Authors: Alban Puech (IBM); Hendrik Hamann; Srihtih Bharadwaj Burra
Contact: Tilman Bockhacker Tilman.bo@gmx.de
Organization: Stony Brook University / Leipzig University

Abstract:
This poster presents my master thesis, which concerns an extension of the GridFM foundation model to forecast the Alternating Current Optimal Power Flow (AC-OPF) problem. The presentation focuses on comparing a traditional two-step baseline against a novel end-to-end deep learning approach.

The classic baseline uses a decoupled process: a predictive model first forecasts future electrical loads, which are then used to solve the AC-OPF equations. In contrast, the proposed GridFM approach uses a single unified model to simultaneously forecast loads and generate the corresponding AC-OPF solutions.

As this is ongoing research, the poster specifically explores the architectural designs required for the temporal extension of GridFM. Multiple methodologies for processing time-series data within the model are explored. The theoretical advantages and structural trade-offs of Temporal Attention heads, Temporal Convolutional Networks (TCN), and Recurrent Neural Network (RNN/LSTM) units will be compared.

Poster 21: APPFL-SIM: Easy and Simple Federated Learning Simulation

Authors: SEOKJU (ADAM) HAHN (Argonne National Laboratory); KIBAEK KIM (Argonne National Laboratory)
Contact: SEOKJU (ADAM) HAHN hahns@anl.gov
Organization: Argonne National Laboratory

Abstract:
We present APPFL-SIM, a simulation-first spinoff of APPFL that enables rapid prototyping of federated learning (FL) algorithms while preserving seamless portability to production-scale, cross-facility deployments. As foundation and grid-scale models increasingly require collaboration across institutions, researchers face a gap between lightweight algorithm PoC and HPC-grade distributed execution. APPFL-SIM addresses this gap through a unified backbone architecture: algorithms developed and validated locally with public FL benchmark datasets can be migrated without modification to APPFL’s production runtimes (MPI, gRPC, Globus Compute), enabling reproducible scaling from laptop experiments to national-lab infrastructure. By treating simulation as reference semantics and enforcing adapter equivalence across runtimes, APPFL-SIM reduces research-to-production burden while maintaining correctness and reproducibility. This design provides a practical pathway for grid-scale foundation model research that demands both rapid iteration and cross-site operational reliability.

Poster 22: Reliability-Aware Control of Distributed Energy Resources using Multi-Source Data Models

Authors: Gejia Zhang, Robert Mieth
Contact: Gejia Zhang gejia.zhang@rutgers.edu
Organization: Rutgers University – New Brunswick

Abstract:
Distributed energy resources offer a control-based option to improve distribution system reliability by ensuring system states that positively impact component failure rates. This option is an attractive complement to otherwise costly and lengthy physical infrastructure upgrades. However, required models that adequately map operational decisions and environmental conditions to system failure risk are lacking because of data unavailability and the fact that distribution system failures remain rare events. This project addresses this gap and proposes a multi-source data model that consistently maps comprehensive weather and system state information to component failure rates. To manage collinearity in the available features, we propose two ensemble tree-based models that systematically identify the most influential features and reduce the dataset’s dimensionality based on each feature’s impact on failure rate estimates. These estimates are embedded within a sequential, non-convex optimization procedure, that dynamically updates operational control decisions. We perform a numerical experiment to demonstrate the cost and reliability benefits that can be achieved through this reliability-aware control approach and to analyze the properties of each proposed estimation model.

Poster 23: Hierarchical Multi-Modal Cyberattack Detection for Distributed Energy Resource Operations Using Large Language Models

Authors: Sunho Jang (Brookhaven National Laboratory), Jorel Austin Abrantes (Clovis Community College) , Meng Yue (Brookhaven National Laboratory)
Contact: Sunho Jang sjang@bnl.gov
Organization: Brookhaven National Laboratory

Abstract:
The increasing digitalization of the power grid expands the cyberattack surface of modern power grids, particularly against coordinated attacks that span cyber and physical domains. In this poster, we propose an LLM-based multi-stage, multi-modal cyberattack detection and root cause analysis framework. In the first stage, we leverage modality-specific lightweight anomaly detectors that operate independently to enable real-time detection. For physical grid measurement data, we propose a novel deep learning-based anomaly detection that uses Pearson-correlation as a feature vector, enabling efficient detection using a compact neural architecture. For network traffic data, we introduce a method that leverages discriminative features selected via correlation analysis with attack labels and evaluated using a reconstruction-based anomaly detection model. These detectors effectively identify single-modality attacks with low computational overhead. On top of these detectors, we introduce a Large Language Model (LLM)-based detection and reasoning layer to detect sophisticated cross-domain attacks that evade unimodal detection. We first convert representations of network packets, grid states, and log events into semantic prompts, allowing the LLM to perform cross-modal consistency analysis and constraint-aware reasoning. By improving LLMs performance using in-context learning, chain-of-thought prompting, and targeted fine-tuning, the LLM captures latent dependencies across modalities, identifies coordinated cyberattacks, and generates interpretable root-cause analyses of cyberattacks.

Poster 24: GPU-to-Grid: Voltage Regulation via GPU Utilization Control

Authors: Zhirui Liang, Jae-Won Chung, Mosharaf Chowdhury, Jiasi Chen, Vladimir Dvorkin
Contact: Zhirui Liang zhirui@umich.edu
Organization: University of Michigan

Abstract:
While the rapid expansion of data centers poses challenges for power grids, it also offers new opportunities as potentially flexible loads. Existing power system research often abstracts data centers as aggregate resources, while computer system research primarily focuses on optimizing GPU energy efficiency and largely ignores the grid impacts of optimized GPU power consumption. To bridge this gap, we develop a GPU-to-Grid framework that couples device-level GPU control with power system objectives. We study distribution-level voltage regulation enabled by flexibility in LLM inference, using batch size as a control knob that trades off the voltage impacts of GPU power consumption against inference latency and token throughput. We first formulate this problem as an optimization problem and then realize it as an online feedback optimization controller that leverages measurements from both the power grid and GPU systems. Our key insight is that reducing GPU power consumption alleviates violations of lower voltage limits, while increasing GPU power mitigates violations near upper voltage limits in distribution systems; this runs counter to the common belief that minimizing GPU power consumption is always beneficial to power grids.

Poster 25: Self-supervised Learning for AC Optimal Power Flow in Distribution Systems with Hard Constraints

Authors: Hoang T. Nguyen (MIT); Shaohui Liu(MIT); Reetam Sen Biswas (GE Vernova); Nurali Virani (GE Vernova); Deepjyoti Deka (MIT); Priya L. Donti (MIT)
Contact: Hoang Nguyen hoangh@mit.edu
Organization: MIT

Abstract:
The proliferation of behind-the-meter distributed energy resources provides a significant opportunity to dynamically operate distribution grids by actively coordinating these resources. This can provide substantial benefits in lowering costs and enabling cleaner operations, but requires solving optimization problems such as AC optimal power flow (AC-OPF) at much greater speed and scale. Self-supervised learning methods have emerged to provide fast solutions for distribution grid AC-OPF, but often struggle to satisfy operational constraints strictly. To address this, we propose a sequential linearized feasibility-seeking (SLFS) algorithm that exploits linear models and the network structure of distribution grids to efficiently enforce operational constraints. Building on SLFS, we develop two complementary learning frameworks. The first, SLFS-FSNet, integrates feasibility-seeking during both training and inference to provide strong feasibility and optimality guarantees. The second, Train-Penalty and Test-Feasibility (TPTF), applies SLFS only at inference, substantially reducing training time while preserving feasibility. To enable efficient end-to-end training, we embed a differentiable fixed-point power flow solver and introduce an M-step Jacobian approximation with provable error bounds. Experiments on 33-, 123-, 300-, and 1000-bus feeders show that the proposed frameworks achieve constraint violations and optimality gaps comparable to a high-quality nonlinear solver while delivering orders-of-magnitude runtime speedups, demonstrating a practical path toward scalable, physics-consistent, and constraint-aware learning for distribution-grid AC-OPF.

Poster 26: A Two-Phase Vertical Federated Learning Strategy for Efficient FDIA Detection

Authors: Junho Jeong; Minseok Ryu (Corresponding Author)
Contact: Junho Jeong jjeong53@asu.edu
Organization: Arizona State University

Abstract:
While cyberattacks on power systems have emerged as a significant threat in recent years, deep learning- based methods have shown great promise in detecting stealthy False Data Injection Attacks (FDIA) that elude classical model-based approaches. However, the expansive scale of modern power grids makes centralized training of these neural networks increasingly impractical due to privacy, storage concerns.

To address these challenges, Horizontal Federated Learning (HFL) has been proposed; however, it is often impractical for power system applications where monitoring devices are spatially distributed across wide-area grids. Such infrastructure naturally results in vertically partitioned data across various entities, creating disjoint feature spaces rather than a uniform one. While Vertical Federated Learning (VFL)-based FDIA detection has been introduced to mitigate these spatial constraints, standard VFL architectures remain inherently communication-intensive. Furthermore, they require abundant labeled datasets, which is unrealistic in practice given the high costs of data labeling in power systems.

To overcome these limitations, we aim to design a VFL training strategy that maximizes both communication and data efficiency. Specifically, our approach introduces a two-phase training framework: an initial pre-training phase utilizing readily available unlabeled data, followed by a fine-tuning phase that incorporates multiple local updates to reduce communication overhead in VFL.

Poster 27: VoltGuard: Alarm-Aware Volt/VAR Control via Preference Learning and Constrained Policy Optimization

Authors: Alaa Selim
Contact: Alaa selim alaa.e.selim@dartmouth.edu
Organization: Dartmouth College

Abstract:
Volt/VAR control in modern distribution feeders is judged not only by voltage outcomes, but also by whether issued commands look normal to utility monitoring systems. In practice, large setpoint jumps, frequent tap/cap switching, or PV curtailment inconsistent with irradiance can trigger alarms and lead to command blocking or conservative fallbacks. This creates a gap: a controller can reduce voltage violations in simulation yet be impractical in the field because its commands are flagged. We propose VoltGuard, an alarm-aware Volt/VAR control framework that accounts for monitoring during optimization. At each step, VoltGuard proposes commands for PV reactive power, PV curtailment, battery active power, regulator taps, and capacitor states. A monitoring module produces an alarm score based on command changes, device limits, consistency with external conditions, and the resulting voltage response; a governor limits or replaces commands that exceed an alarm threshold. Because “acceptable” behavior is hard to express as a hand-tuned reward, VoltGuard learns it from comparisons. Using an AC power-flow simulator (OpenDSS), we generate pairs of candidate command outcomes and summarize each outcome using voltage quality, violation magnitude, switching activity, and command smoothness. A judge (LLM and/or limited expert labels) selects the more acceptable outcome, and we train a reward model from these preferences. Finally, we train a policy with constrained reinforcement learning to maximize the learned reward while keeping the average alarm score below a chosen budget, yielding an explicit voltage–alarm tradeoff.

Poster 28: Minimizing Bid Cost Recovery under Uniform Pricing with Energy Storage

Authors: Yaxuan Yu; Jingguan Liu; Cong Chen
Contact: Yaxuan Yu yaxuan.yu.th@dartmouth.edu
Organization: Dartmouth College

Abstract:
We study uniform pricing for real-time rolling-window economic dispatch in power systems with energy storage resources (ESRs). In multi-interval dispatch, intertemporal state-of-charge (SOC) dynamics couple decisions across time, and settlement based on standard locational marginal pricing (LMP) can lead to revenue inadequacy and out-of-market bid cost recovery (BCR) uplifts. We propose an online uniform-pricing framework that selects a single energy price at each interval by minimizing the total BCR, subject to revenue adequacy for cleared participants. We further introduce a regularized variant that balances BCR reduction against proximity to a benchmark marginal price, providing explicit control over the trade-off between uplift reduction and price consistency. We establish an impossibility result showing that no single uniform price can eliminate BCR for ESRs under rolling-window dispatch with SOC constraints. Monte Carlo rolling-window simulations with multiple ESRs under forecast uncertainty compare the proposed rules with LMP, temporal LMP (TLMP), and max temporal LMP (MTLMP). Results demonstrate that the proposed uniform-pricing rules substantially reduce total BCR and demand payments relative to LMP, while yielding lower generator profits and driving merchandising surplus to zero under uniform-price settlement.

Poster 29: Enabling Reliability-Aware Capacity Expansion Planning with Grid Foundation Models

Authors: Sunash Sharma (UCB, ANL); Jonghwan Kwon (ANL); Kibaek Kim (ANL); Duncan Callaway (UCB)
Contact: Sunash Sharma sunash.sharma@anl.gov
Organization: University of California, Berkeley; Argonne National Lab

Abstract:
Grid foundation models (GridFMs) have shown promise solving a wide set of operational problems in power systems, but the application of these models to capacity expansion planning (CEP) remains underexplored. We present ongoing work on the integration of a GridFM into CEP workflows to yield low-cost, reliable investment decisions. In particular, we aim to use GridFM to perform fast prediction of traditionally computationally burdensome reliability assessment (RA) outputs, enabling broad but tractable inclusion of reliability criteria into capacity expansion planning.

We are generating RA data for a wide variety of systems, with variation over topology, loading, resource availability, and outages using A-LEAF, a simulation framework for power system operations and planning at Argonne. Using this data, we will pursue an initial fine-tuning of an existing GridFM to predict unserved energy for RA scenarios on a subset of systems and scenarios, and will evaluate out-of-sample performance and generalization in subsequent experiments. Finally, we propose methods to integrate a fine-tuned GridFM within CEP to act as a surrogate for RA. These methods include direct embedding of the surrogate within CEP, using the surrogate in iterative solution processes that can include RA information (such as Benders’ decomposition), and using GridFM to guide high-risk scenario generation. Here, we present the data generation pipeline and a small-scale fine-tuning proof of concept, and we provide a roadmap for RA surrogate integration into CEP.

Poster 30: VoltGuard: Alarm-Aware Volt/VAR Control via Preference Learning and Constrained Policy Optimization

Authors: Alaa Selim, Junbo Zhao
Contact: Alaa Selim alaa.e.selim@dartmouth.edu
Organization: Dartmouth College

Abstract:
A Volt or VAR controller can keep voltages within limits in simulation and still fail in practice when utility monitoring flags its commands as abnormal and triggers blocking or safe fallback actions. We present VoltGuard, a reinforcement learning with AI feedback (RLAIF) framework for Volt or VAR control (VVC) that treats this security stage as part of the control problem rather than an afterthought. VoltGuard operates on an AC power flow simulator (OpenDSS) and issues coordinated setpoints for photovoltaic inverters, battery energy storage systems, regulator taps, and capacitor banks. A monitoring module computes a flag risk score from command size and rate of change, device limit compliance, consistency with external conditions such as irradiance, and the resulting voltage response; a governor then screens high risk commands. VoltGuard learns an operational acceptability reward from pairwise comparisons. Under the same operating conditions, it simulates two candidate command outcomes, summarizes each outcome by voltage quality, violation severity, and switching activity, and trains a preference model that assigns higher reward to the more acceptable outcome. A constrained policy optimizer then improves voltage regulation while keeping average flag risk below a chosen budget, producing a clear tradeoff between voltage performance and alarm exposure. We evaluate VoltGuard on the Iowa distribution test system and show that it provides actionable intuition for daily VVC operations by reducing voltage violations while avoiding the abrupt, high frequency actuation patterns that monitoring systems typically flag.

Poster 31: Reducing Renewable Uncertainty: A Roadmap Using Storage, Forecasting and Control

Authors: Dr. Bolun Xu, Columbia University
Contact: Yang Du yang.du@jcu.edu.au
Organization: James Cook University

Abstract:
High penetrations of variable renewable energy (VRE) are reshaping how power systems manage variability and uncertainty. In the Australian National Electricity Market (NEM), large-scale wind and solar operate as semi-scheduled resources, with forecast errors and dispatch deviations managed ex post through frequency control ancillary services (FCAS) and recovered via Frequency Performance Payments (FPP). This “cause-and-fix” paradigm assumes external reserves can absorb VRE uncertainty at acceptable cost, but recent experience—including the October 2025 frequency event—suggests it becomes increasingly fragile as VRE shares grow. In this work, we examine how the NEM translates uncertainty into dispatch error, FCAS procurement, and FPP allocations. We then synthesize practical pathways to firm renewable output, including PV overbuild with proactive curtailment, improved forecasting and controls, and energy storage. Building on these insights, we discuss how firm generation can be applied under different market designs. The framework shifts emphasis from correcting uncertainty after it enters the system to removing most uncertainty before dispatch, supporting secure high-renewables operation in the NEM.

Poster 32: Graph Neural Networks for Power System State Estimation and PMU‑Based Security Assessment

Authors: Sujit Tripathy (EPRI); Vikas Singhvi (EPRI); Mahendra Patel (EPRI); Tuna Yildiz (Northeastern University); Ali Abur (Northeastern University); Evangelos Farantatos (EPRI); Nanpeng Yu (UCR)
Contact: Vikas Singhvi vsinghvi@epri.com
Organization: EPRI

Abstract:
Graph Neural Networks (GNNs) offer a natural framework for learning from power system topology and measurements. This work presents two GNN-based applications for large-scale power systems. First, a graph attention network (GAT)–based model is used for missing data imputation in state estimation, where voltage magnitude and angle at unobservable buses are estimated using network topology and observable measurements. The estimated states are used to generate synthetic measurements with low weights in a weighted state estimation framework, improving system state awareness under partial observability. Second, a PMU-based dynamic security assessment framework is developed using graph convolutional and graph attention networks to predict frequency, voltage, and rotor angle health indices under contingencies. Results on a large test system demonstrate that GNN models outperform conventional multilayer perceptrons, with GAT models achieving the highest prediction accuracy and computational efficiency. The work highlights the potential of GNNs for scalable, data-driven monitoring and security assessment of modern power grids.

Poster 33: GridSearch – Data Center Hosting Maps with GridFM

Authors: Alban Puech (IBM); Hendrik Hamann; Srihtih Bharadwaj Burra
Contact: Hendrik Hamann hendrik.hamann@stonybrook.edu

Abstract:
Interconnecting new loads and generation to the electric grid is a complex and costly process, which often stretches over many years, frequently requiring major additional infrastructure such as substations, switch gear, transformers, or transmission lines, which in turn will determine additional project delay and cost. Here we propose a new method (GridSearch) for accelerating this interconnection process, which is based on the idea of throughout prescreening of many potential interconnection locations which are likely to have the least amount of impact on the electric grid before embarking on the cumbersome interconnection process. GridSearch utilizes GridFM), which can enable faster grid simulations while being accurate and topology robust. The results from a GridSearch analysis can be visualized as maps (interconnection complexity maps) depicting the interconnection complexity, which is assessed by GridSearch through a contingency analysis after the new hypothetical load/generation is connected.