Abstract

Power system transient event analysis is essential for secure and reliable grid operation and increasingly requires the joint use of heterogeneous information. Although recent data-driven methods have explored multimodal learning for transient analysis, effectively integrating these modalities remains challenging: fixed fusion strategies may not generalize across tasks, and physical topology alone may not capture data-driven inter-bus dependencies. To address these challenges, we propose a novel Mixture-of-Fusion-Experts (MoFE) framework for adaptive multimodal transient event analysis. The framework incorporates three fusion experts operating at early, intermediate, and late fusion stages, together with a task-aware router that adaptively selects the appropriate fusion granularity for different tasks. To enhance the representation of network relationships, we construct a causal topology graph that augments the physical network with data-driven inter-bus dependencies. We further introduce a memory bank that retrieves representations of relevant historical operating conditions to enrich the current event representation and improve generalization across varying system conditions. Extensive experiments demonstrate that the proposed MoFE framework consistently outperforms state-of-the-art baselines across multiple transient analysis objectives.

Abstract

The AI-EFFECT Testing and Experimentation Facility (TEF) addresses the challenges of systematically testing, validating, and deploying AI in electricity systems, which is relevant for emerging paradigms such as foundation models. This poster presents the Portuguese AI-EFFECT node, a specialised testing and experimentation environment targeting residential consumers and low-voltage (LV) electricity grids, with a focus on its design and implementation.

The node integrates software, hardware, energy data and data spaces, AI services, and computational infrastructure to support the development and evaluation of AI-driven applications. Its use cases cover AI-enabled energy management and energy sharing, distributed energy resource scheduling and control, load and renewable generation forecasting, and LV distribution-network monitoring. AI providers can evaluate their solutions across different levels of experimentation, ranging from simulation to laboratory and living-lab environments, supported by the TEF services available within the node.

The poster presents the node’s use cases, technical architecture, and integration with the AI-EFFECT platform, highlighting its data, computational, orchestration, and testing capabilities. Preliminary implementation results demonstrate the potential of this integrated environment to facilitate the systematic validation of AI solutions and accelerate the deployment of next-generation AI applications in electricity systems.

Abstract

The increasing penetration of converter-interfaced resources (CIRs) is profoundly changing power-system dynamics, making stability increasingly dependent on converter control settings, operating conditions, and interactions among multiple devices. This calls for analysis tools capable of efficiently exploring large multidimensional parameter spaces while retaining a system-level representation of grid dynamics. This work presents an AI-assisted framework for stability and sensitivity analysis of converter-dominated power systems. Building on a detailed small-signal modeling approach, synchronous generators, converter-interfaced resources, network components, and their control systems are integrated into a unified state-space representation. Eigenvalue analysis over multiple operating conditions is combined with machine-learning techniques to efficiently characterize the relationship between converter parameters and system stability. In particular, adaptive sampling and probabilistic classification are exploited to identify stability regions while concentrating computationally expensive evaluations close to critical boundaries. The proposed framework extends the concept of stability manifolds toward a broader data-driven sensitivity analysis, aimed at identifying the converter parameters and parameter interactions that most significantly affect system stability and dynamic performance. Further developments will investigate alternative AI techniques, different grid-following and grid-forming control architectures, and uncertainty in converter parameters and equivalent models. The resulting methodology is intended to provide scalable tools for stability assessment and converter integration in future power systems.

Abstract

Load forecasting will be among the first tasks asked of a grid foundation model, and the ECL dataset (hourly demand from 321 customers) is its de facto benchmark. We evaluated 17 forecasting architectures, spanning classical, MLP, convolutional and Transformer families, under two evaluation configurations at horizons of 96 to 720 hours.

Three findings emerge. First, the leaders are nearly tied: at 96 hours, N-BEATS (MSE 0.145), N-HiTS (0.147) and iTransformer (0.148) differ by about 2%, while ARIMA errs roughly eight times more. Second, rankings depend on horizon: N-HiTS error grows 65% from 96 to 720 hours against 26% for FEDformer, which overtakes it at the longest horizon. Third, configuration matters as much as architecture: FEDformer’s MSE moves from 0.183 to 0.193 between setups, a shift larger than the gap among the top three models.

Rather than a bigger model, we argue the field needs a better test. ECL has saturated as a discriminator, and it lacks the topology, weather and distributed-generation signals a GridFM is built to exploit. We propose a minimal reporting standard (fixed look-back, multiple seeds, horizon-wise curves) so that reported gains exceed evaluation noise. We also outline a grid-aware successor benchmark pairing feeder-level load with topology and weather covariates.

Abstract

Distribution grids were traditionally designed for unidirectional power flow, from substations to customers. As increasing amounts of wind and PV generation connect to medium-voltage feeders, bidirectional power flows can lead to line overloading and bus-voltage violations.

We treat voltage control as a sequential decision problem. At five-minute intervals through the day, a controller sets the on-load tap changer of the HV/MV transformer and the active and reactive power of each controllable generator. The goal is to keep all buses within the ±5 % operational band and the ±10 % EN 50160 limit while curtailing as little active power as possible. These two aims conflict, as the as stricter voltage regulation can require increased active power curtailment.

We train a deep reinforcement learning agent on an AC power flow digital twin. It controls the tap and 16 wind and PV units of the SimBench rural MV grid, trained on a year of historic data. We compare it with no control, local tap control, and an oracle OPF with perfect foresight.

On 56 unseen test days, the agent removes nearly all voltage violations and never breaks the regulatory limit. Its performance approaches that of the oracle with perfect foresight, at slightly higher curtailment, while computing each decision in milliseconds.

Abstract

Learned surrogates for power-flow (PF) computation are usually motivated by speed, compared against a single CPU-based baseline. We show that this comparison is fragile: across 7 open-source CPU-based solvers evaluated on identical inputs, throughput itself varies by roughly 2.5–3 orders of magnitude, so the choice of baseline alone can flip the conclusion. We build a benchmark that instead spans 9 physics-based solvers — 7 CPU-based and 2 GPU-based, including a new batched GPU solver, gpu-0DCD — and 3 published ML surrogates, evaluated under a shared, independently implemented KCL-mismatch check across 15 grids from 14 to 70,000 buses (3.5 orders of magnitude). On the PF task itself, a sufficiently parallelized CPU solver or gpu-0DCD matches or exceeds the throughput of the ML surrogates we test while remaining orders of magnitude more accurate; we find no evidence that these GNN-style surrogates are yet competitive for PF specifically, though prior work reports genuine gains on related but more combinatorial tasks such as optimal power flow. Beyond this specific finding, we hope the benchmark suite — code and reference evaluator released at an anonymized repository — helps close a broader reproducibility gap in this space, where batched GPU-based PF solvers are rarely open-sourced or evaluated against a common set of baselines.

Abstract

The increasing application of data-driven methods and foundation models in power systems creates a growing demand for large-scale, high-quality, and representative datasets. However, real-world power system measurements are inherently limited, as data collection typically covers only a few years of operation. Moreover, privacy, security, and infrastructure-related constraints can restrict the sharing of measurement data.

This work investigates measurement-informed synthetic data generation to address these limitations. Real-world measurements collected at the Living Lab Energy Campus (LLEC) at Forschungszentrum Jülich are used as the basis for generating synthetic power system time series that reproduce relevant characteristics of the measured data without directly sharing the underlying measurements, providing data that are realistic but not real. The aim is to generate privacy-preserving, shareable, and representative datasets that extend beyond the limited snapshot of system behavior captured by finite measurement periods.

Two complementary approaches are investigated: data-driven generative models, including pretrained Grid Foundation Models adapted using site-specific measurements, and classical power system simulation models combined with measured data. Both approaches integrate real-world system behavior with scalable data generation while incorporating different degrees of learned and physical knowledge. The resulting synthetic datasets aim to enable privacy-conscious data sharing and provide representative data for the development and evaluation of data-driven power system applications.

Abstract

Grid foundation models (GridFMs) are pretrained at scale on power system data, enabling fine-tuning for diverse downstream tasks and generalization to networks unseen during training.Research on GridFMs has focused mainly on predictive performance, but critical infrastructure also demands explainability: operators and regulators must be able to verify that a model’s outputs are consistent with grid physics. We aim to fill this gap as follows. First, we map the capabilities of existing GridFMs such as GridFMv0.2, GridSFM, and MxGPS, to the demands they place on explainability. We keep this mapping adaptable to accompany future, more expressive models. Second, we introduce ExplainKit, which offers explanations, metrics, physics-derived reference masks, and tools for what-if (counterfactual) analysis and visualization. ExplainKit generalizes explainers designed for message-passing graph neural networks to the graph transformer architectures of current GridFMs. Initial results on power flow, comparing against ground-truth sensitivities derived from the power-flow equations, show that even in-distribution these models capture the physical structure only partially, so accuracy does not guarantee faithfulness. We further diagnose known shortcomings, such as poor generalization to unseen and larger grids, and relate them to modeling choices ranging from data generation and feature design to architecture. With ExplainKit, we also support explainability efforts for weather and other scientific (GNN) foundation models.

Abstract

Driven by the transition towards a climate-neutral energy system, accurate energy time series forecasting is critical for planning and operations. Yet, it remains a dataset-specific task, requiring comprehensive training data, limiting scalability, and resulting in high model development and maintenance effort. Recently, foundation models aiming to learn generalizable patterns via extensive pretraining have shown strong performance in multiple prediction tasks. Despite their success and strong potential in energy forecasting, a systematic, use-case-differentiated evaluation is still missing. We address this gap by presenting the Foundation Models in Energy Time Series Forecasting (FETS) benchmark. We (1) provide a structured overview of energy forecasting use cases along three main dimensions, i.e., stakeholders, attributes, and data categories, (2) curate 54 datasets across 9 data categories, guided by typical stakeholder interests, and (3) benchmark foundation models against task-specific machine learning across different forecasting settings. In our benchmark study, covariate-informed zero-shot foundation models perform best in aggregate, with Chronos-2 attaining the lowest overall median NRMSE (0.472), closely followed by TiRex-2 (0.474). Both perform better than XGBoost (0.611) and random forest (0.696), although they were trained task-specifically on the full historic target data. Further analysis reveals a strong correlation between predictive performance and spectral entropy. Performance saturates beyond a certain context length and improves with aggregation level, e.g., for national load, district heating, and power grid data. Overall, with the lowest median error, limited data requirements, and low inference and hardware demands, foundation models reduce development and maintenance effort, emerging as scalable and generalizable energy forecasting solutions.

Abstract

The integration of distributed energy resources (DERs) is increasing rapidly to support the energy transition. However, their diverse and variable adoption patterns pose challenges for distribution system operators (DSOs), requiring them to evaluate large sets of scenarios under uncertainty. Advances in generative AI enable the generation of diverse and realistic scenarios conditioned on future planning assumptions (e.g., demand growth). Although these scenarios better represent future uncertainty, conventional planning approaches are overly conservative and limit DER integration. However, from a planning perspective, an explicitly quantified risk should be accepted because violations can be managed through operational strategies (e.g., flexibility and curtailment). This allows DSOs to integrate more DERs while remaining informed about the associated risks. This work presents a risk-based framework for PV-rich distribution system planning. A conditional flow-based model is used to generate load profiles conditioned on annual energy consumption growth. The load profiles are combined with stochastic irradiance scenarios to represent uncertainty over the planning horizon. The generated scenarios are evaluated using Monte Carlo-based probabilistic power flow simulations, while voltage violations are assessed through three metrics: intensity, duration, and frequency, capturing severity, persistence, and likelihood, respectively. Results show that allowing a 5% risk level increases PV hosting capacity by 18% for a 15 min violation duration.

Abstract

AI offers a promising path to accelerate EMT simulation, but existing AI-based EMT models are often task-specific and difficult to generalize. This work develops a grid foundation model for EMT simulation (GridFM-EMT) that learns reusable EMT representations from diverse trajectories and adapts to unseen scenarios through lightweight fine-tuning. GridFM-EMT shifts EMT learning from isolated surrogates toward a pretrained, reusable, and generalizable simulator.

Abstract

Foundation models for the electric grid are usually pretrained by masked reconstruction: part of a grid snapshot is hidden and the model learns to fill in the missing values. Other domains such as vision and video take a different route, joint-embedding predictive architectures (JEPA), which predict an abstract representation of the hidden part instead of its raw values. These methods have been reported to gather more informative features and to be more robust to noisy inputs. We bring this idea to power systems with Grid-JEPA, a compact graph model pretrained on synthetic distribution grids. We then adapt it to two core operational tasks, power flow and state estimation. We ask whether latent prediction learns more than reconstruction, whether it beats an existing grid foundation model, whether it transfers from medium- to low-voltage grids, and whether it adapts quickly to a new grid. Grid-JEPA captures more information about the grid’s operating state and is the strongest learned model for power flow on unseen grids, while being about twenty times smaller than GridFM. Pretrained only on medium-voltage feeders, it gives the best state estimates on public low-voltage grids, and it needs few labelled examples to adapt. Latent prediction is a promising pretraining objective for grid foundation models.

Abstract

Foundation models for power systems should support multiple protection tasks, but strong fault-classification performance does not establish that a representation also supports fault location. As a step toward the GridProtect roadmap, we examine what frozen waveform representations make accessible across these tasks. On matched line-fault populations from three simulated grids, we compare random and self-supervised CNN encoders with linear and bounded nonlinear readouts, supervised end-to-end training, and a frozen pretrained time-series foundation model. Each comparison uses five seeds per grid.

For fault classification, nonlinear readouts improve the self-supervised encoder on all three grids by 0.033 – 0.061 supported macro-F1; pretraining provides a smaller and less consistent gain over matched random features. For fault location, the nonlinear readout has no consistent benefit on CIGRE MV but reduces mean absolute error by 5.76 and 3.40 percentage points on TestGrid110kV and IEEE39. Supervised adaptation improves fault location on every grid. Preserving channel identity substantially improves the pretrained model’s classification readout, yet its matched-capacity representation remains below even the random CNN for classification in every observed seed.

These validation results show how task, grid, readout capacity, and channel structure shape claims about representation sufficiency for protection.

Abstract

Voltages depend on the whole network, but corrective graph neural networks for power flow exchange information locally. We add a hierarchical exchange to GENCO’s corrective network: after five local layers, bus states are restricted twice onto graphs of anchor buses, combined by attention over all top-level anchors, and prolonged back, in place of two of the twelve local layers. Two reductions share the same anchors: Kron reduction, whose weights follow from the grid’s admittance, and a graph-partition quotient. Each variant is one model of about 20 M parameters shared by three grids (500, 2,000 and 4,917 buses), trained for 200 epochs with three seeds, about 100 times fewer updates per grid than GENCO’s reference training. On 600 newly generated scenarios on these grids, Kron reaches a macro voltage error of 0.851 ± 0.110, 51.3% below a per-bus mean fitted on training solutions (1.747), and beats that reference in 98.5% of scenarios. Both hierarchies beat it on every grid in every seed; Kron is 31.0% below the quotient (1.235 ± 0.225). The unmodified GENCO Base architecture, trained the same way, reaches 5.660 ± 0.899 and stays above the reference. On two unseen topologies, the models do not yet beat fitted references in calibrated transfer; this is our next objective.

Abstract

Learning by examples: power grid steady-state analysis with a tabular foundation model

Foundation models have shown strong generalization across scientific domains, but their use in power systems remains largely task-specific. We investigate whether a pretrained tabular foundation model (TFM), TabPFN v3, can serve as a general-purpose neural surrogate for power grid steady-state analysis. We evaluate the TFM on three tasks: power flow (PF) prediction, N-k contingency analysis (NCA), and voltage stability margin (VSM) prediction. We performed experiments on 118- and 2000-bus systems, comparing the TFM against state-of-the-art task-specific GNNs. Without retraining or fine-tuning, the TFM matches or outperforms the best baseline across the three tasks. For 118-bus PF prediction, it achieves normalised RMSEs of 1×10^(-3)pu for voltage magnitudes and 0.30° for voltage angles, outperforming all baselines. In zero-shot N-1 and close-to-infeasible PF cases, it also outperforms the DCPF approximation. For millions of unseen N-2 contingencies, the TFM achieves AUCs of 0.92 and 0.94 for branch and voltage violation screening, respectively. The TFM also performs strongly on VSM prediction for the 2000-bus system with limited labelled data. These results demonstrate the potential of pretrained TFMs as a neural surrogate for power grid analysis, despite lacking power-system-specific inductive biases or pretraining data.

Abstract

Unit Commitment is a power system scheduling problem that can be formulated as a mixed-integer optimisation, requiring thousands of binary on/off decisions for generators to be resolved over each scheduling horizon. This poster presents a machine learning approach that accelerates this process by predicting which decisions can be safely fixed before optimisation begins, allowing the solver to focus on the decisions that matter most. Rather than replacing the optimiser, the approach combines machine learning with conventional optimisation, using prediction confidence together with the potential impact of errors to decide which decisions to fix. The method formed the basis of our first-place solution in EPRI’s AI-ccelerating Unit Commitment Competition, achieving an average speed-up of more than 50× across diverse operating conditions. The results highlight how machine learning can reduce the computational burden of optimisation while retaining the solver’s ability to handle uncertain or consequential decisions.

Abstract

Distribution system operators and aggregators increasingly need accurate voltage estimates under rising photovoltaic (PV) penetration, yet detailed feeder models are often unavailable, costly, or outdated. Purely data-driven neural networks trained on smart-meter data interpolate well but generalize poorly to unseen operating regimes. We present a physics-informed neural network (PINN) whose physical constraint is itself derived from data: a voltage-sensitivity loss that enforces the per-customer ΔV/ΔP relationship estimated from historical smart-meter measurements. Unlike existing physics-informed approaches that rely on Jacobian matrices or network parameters, the method requires no topology or impedance information, enabling truly model-free deployment. On a realistic Japanese distribution feeder simulated in OpenDSS, the PINN matches a conventional neural network in interpolation and clearly outperforms it in extrapolation: trained only on no-PV data, it reproduces midday voltage rise at up to 100% PV penetration, reducing RMSE by about 70%. Inference is roughly 20× (CPU) and 230× (GPU) faster than power-flow calculation on a 10,867-bus feeder, with negligible overhead. These results show that lightweight, data-derived physical priors can make grid AI models more trustworthy under distribution shift.

Abstract

MxGPS (Multiplex GPS) is a graph transformer for multi-task power grid analysis. A shared encoder embeds bus features and structural encodings, then feeds two parallel GPS branches, one for state estimation (SSE) and one for AC power flow (PF), each conditioned by a learnable task token. Each GPS layer combines local GatedGCN message passing with global edge-biased attention. After masked-reconstruction pre-training, both branches are fine-tuned jointly with a power-balance physics loss and a voltage-boundary penalty. Across three cross-validation folds on 14–300-bus IEEE systems, single-task models degraded by 190–1400% on unseen grids, while MxGPS degraded by 39% and produced no voltage-limit violations on any unseen topology, with 12× fewer parameters than the GridFM baseline. The trade-off is higher in-distribution error and compressed voltage predictions. Our team is now focusing on (i) replacing the homogeneous GPS backbone with heterogeneous (e.g., HGNS); (ii) hold out intermediate-sized grids to separate topology transfer from size extrapolation; (iii) scale to grids with thousands of buses using sub-quadratic attention; (iv) add tasks such as optimal power flow and N-1 contingency analysis, with multi-seed replication.

Abstract

Renewable and storage operators in ERCOT settle at individual nodes, where the day-ahead-to-real-time spread (DART) can swing by more than $1,000/MWh in a single hour through congestion, curtailment, and local scarcity that hub-level forecasts miss. We present a single deep probabilistic model that forecasts the full hourly DART distribution across roughly 700 ERCOT settlement points: hubs, load zones, and the resource nodes of standalone batteries, utility-scale solar, wind, and solar-plus-storage hybrids. Trained jointly across the market, it learns transferable per-location representations, so data-scarce and newly commissioned nodes inherit structure from the whole grid. A scarcity-aware component targets extreme hours, prediction intervals are calibrated, and every forecast is delivered with its driving features. We adapt the model to ERCOT’s December 2025 real-time co-optimization redesign (RTC+B), a market-structure break. During Winter Storm Fern, the first major scarcity event under RTC+B, the model attained 100% direction accuracy and 100% P99 interval coverage across all nodes on all 54 extreme hours (|DART| > $500/MWh). Over the full storm, direction accuracy was 86-93% versus roughly 60% for prior-day same-hour persistence, node MAE was $38-52/MWh against DART reaching $1,100-1,800/MWh, and P95 and P99 coverage held near nominal at 91.7-95.8% and 97.6-100%. One market-wide model thus delivers calibrated, explainable node-level DART forecasts under ERCOT’s new design.

Abstract

Intraday electricity markets enable participants to adjust energy positions close to delivery, with price forecasts necessary to support trading and the scheduling of flexible electricity resources as well as increasingly responsive consumers. Forecasting model specifications have progressed from using macro-features, such as renewable generation and load, to the micro-features of continuous orderbooks. A recent advanced deep learning model, OrderFusion, explicitly models micro-level buy–sell orderbook interactions. However, despite its superior comparative forecasting performance, it considers only one delivery product at a time, ignoring neighboring-product information. Moreover, when forecasting aggregated price indices such as the liquid German ID3, ID2, and ID1 products, it omits the information contained in price trajectories. In contrast, pretrained time-series foundation models have shown success in financial, renewable-energy, and day-ahead electricity price forecasting. However, their performance on intraday orderbook data remains an open research question of considerable practical importance. In this paper, we propose OrderFusion+, an open-source deep learning model that combines historical orders from the target and neighboring delivery products to forecast probabilistic buy–sell price trajectories. We benchmark OrderFusion+ against forecasting baselines and pretrained foundation models, and investigate dynamic market conditions through the designed dynamic masking mechanism, revealing insights into market efficiency. The implementation and forecasts can be found at: https://runyao-yu.com/OrderFusion/.

Abstract

Time-series foundation models such as Chronos-2 promise strong zero-shot forecasting performance, but it remains unclear whether lightweight adaptation is worth its cost under realistic, operationally-constrained conditions such as demand-response (DR) events. We study this question on EXPESIGNO, a residential smart-meter dataset (175 households, 30-minute resolution, 28 DR events with green/orange alerts, 2019–2021), comparing Chronos-2 zero-shot forecasts against parameter-efficient fine-tuning (LoRA applied to the Q/K/V/O projections of a frozen backbone). Our best configuration (LoRA, r=16, α=32) reduces NRMSE by 3.5% and 2.7% on two household segments, with gains reaching 8.0% on a winter holdout, consistent with a thermosensitivity-driven hypothesis for when adaptation helps most. We also observe a “zero-shot ceiling” effect, where further adaptation yields diminishing returns beyond a certain point. These results offer early evidence on where and how much parameter-efficient adaptation of grid-oriented foundation models pays off under real flexibility/DR conditions, rather than idealized forecasting settings, and contribute to the ongoing discussion on adaptability and data efficiency for Grid Foundation Models.

Abstract

Recent neural power-flow solvers, including emerging foundation models, achieve accurate voltage predictions, yet such accuracy does not necessarily imply physically consistent solutions. Even small complex voltage errors can yield large AC power-balance residuals. We study this accuracy–consistency gap across gridfm-graphkit, LUMINA, GridSFM, and PIGNN-GC on realistic 2224-bus Great Britain network (GBnetwork) scenarios, with cross-grid evaluation over 31 systems.

Using a singular value decomposition (SVD) basis fitted to training AC power-flow solutions, we find that neural prediction errors contain substantial components outside the dominant solution subspace. To address this mismatch, calibrated solution-subspace projection (CSP) suppresses off-subspace prediction components after train-only bias calibration, reducing Mean PB by 40.5%, 68.9%, 37.8%, and 67.0% for gridfm-graphkit, LUMINA, GridSFM, and PIGNN-GC, respectively, relative to calibrated predictions, while improving voltage-magnitude accuracy in all four models. These results identify output-error geometry as an important factor in physics-consistent neural AC power flow.

In ongoing work, we apply CSP inside a rolling-horizon congestion-management loop, where a particle-swarm search issues thousands of AC power-flow evaluations per congested step and approximation error propagates into the dispatch decision.

Code: github.com/Kimchangheon/neural-acpf-error-geometry

Abstract

Synchronized, large-scale computing workloads of datacenters can introduce rapid oscillations affecting the whole power system, creating a growing need for more adaptive support. Conventional controllers have limited ability to respond to changing disturbances, while predictive and learning-based approaches often depend on detailed network models, online optimization, extensive telemetry, or learned value critics. This paper proposes PowerWM, a control framework that learns the nonlinear grid dynamics from historical trajectories, and leverages this model to coordinate the active and reactive power of a co-located battery energy storage system (BESS). Using the learned world model, PowerWM generates imagined multistep rollouts to directly optimize the control policy. System-wide trajectories are required only during world model training, while deployment retains the compact learned controller that relies on local workload and BESS measurements. Across extensive evaluations, PowerWM achieves the lowest voltage and frequency oscillations among the controllers compared on both the 4-bus benchmark and the larger Nordic system, demonstrating scalability across network sizes. PowerWM trained on square waves workloads generalizes without retraining to unseen, realistic measured datacenter oscillations, while sub 2 ms inference enables fast real-time BESS control. These results establish PowerWM as a promising path toward adaptive, locally deployable controllers that can leverage power system world models.

Abstract

Low voltage distribution grids exhibit substantial structural and operational heterogeneity, reflecting differences in topology, phase imbalance, customer density, distributed energy resource integration, and local demand and generation patterns. As these systems become more complex, AI and data driven methods are increasingly explored for analysis, estimation, planning, and operation. Their development requires diverse network datasets that support generalization across heterogeneous grid topologies. However, limited access to real LV network models motivates synthetic grids, whose quality should be assessed by the structural diversity they represent rather than by dataset size alone. In this work, 179 publicly available real LV distribution networks are characterized using structural descriptors and unsupervised learning. The resulting clusters define distinct regions of the LV topology space and provide a data driven reference for generating synthetic populations with different structural characteristics. The 3,000 synthetic LV grids generated are evaluated using four complementary topological diversity indicators: global dispersion, local redundancy, structural coverage, and effective rank. Their evolution with population size is analysed within each cluster to quantify how structural information expands and when additional networks provide diminishing diversity gains. The framework enables quantitative assessment of synthetic LV grid topological diversity, supporting more informative datasets for grid foundation models.

Abstract

The N-1 criterion is an operational rule that requires all possible single component outages – referred to as N-1 contingencies – to be tested and simulated for point forecast scenarios. Unfortunately, this paradigm falls short in view of the growing uncertainty over power injections, making the establishment of a probabilistic security assessment more needed than ever. Multiple research avenues have been explored over the past decades, ranging from simulation time reduction to fast screening heuristics, either relying on simplifying assumptions or learning from historical data. The present work stands out from previous contributions by framing the contingency screening as an edge-level classification problem using the open-source module EnerGNN. More specifically, a Graph Neural Network model is trained to predict the security status for a postulated subset of credible contingencies for a given operating condition in a single forward pass. The proposed methodology is built on real-life subsets containing mostly N-1 contingencies while naturally extending to subsets including contingencies of any order. Here, we focus specifically on detecting contingencies that lead to current violations, which we define as insecure. Validation on real-life, full-scale data from the French high-voltage (HV) and extra-high-voltage (EHV) transmission system showcases strong performances. Under typical operating conditions, less than 1% of N-1 contingencies are truly insecure. Yet, by flagging an average of 1.98% of all N-1 contingencies as potentially insecure, our model successfully detects 98.05% of the actual insecure N-1 contingencies. This ensures fewer than one undetected insecure N-1 contingency (0.43) per operating condition, while requiring less than 50 CPU ms per screening.

Abstract

To address the need for generalizable representation in foundation grid models, we propose a heterogeneous GraphSAGE framework that learns tokens for grid elements representation by combining temporal behaviour and physical attributes with network context obtained through edge message passing. The model is jointly trained with multiple objectives to encourage general-purpose representations, and is evaluated on multi-horizon forecasting and state estimation. Across two ACTIVSg grids, the model achieves in most cases performance comparable to task specific baselines while supporting zero-shot transfer to an unseen grid. The results suggest that the learned representations can generalize across both tasks and topologies.

Abstract

Transmission system operators inspecting high-voltage insulator strings using inference using drone capture inference in remote locations need a local LLM assistant that is useful only if its answers are hallucination-free, drawn from operator-held documents and never exceeding what a photograph shows. The assistant presented here is built accordingly, a detector turning photographs into fixed-text findings, retrieval covering 3,780 IEC, CIGRE and IEEE passages and 127 engineer-reviewed question-answer pairs, and code-enforced hallucination guards checking every answer. Because everything runs within the memory restrictions of an edge device such as the 2 GB Arduino UNO Q, the language model must fit the device, so a weight-sharing supernetwork is trained once from Llama 3.2 1B and eleven subnetworks varying in size are extracted and distilled from a Qwen3.5-27B teacher, yielding a deployable model family. Which network is selected is decided by a selection rule, evaluation set and GPT-OSS:20B judge, ranking candidates by answer quality under a hallucination guard gate and memory restrictions. All subnetworks keep guard violations at or below 2.8% and answer the field questions on the device without failure at 2.8 W, the service peaking at 1108 to 1258 MB as quality rises along the memory-quality Pareto frontier from 0.73 to 0.91 and the median 704-token answer lengthens from 79 to 134 s. The shipped network is the fastest of the three tied on quality within 1200 MB, at 0.85, 94 s per answer and 1172 MB.

Abstract

Foundation models for power system planning and operation require large amounts of high-quality training data. This is particularly challenging for dynamic stability assessment, where historical data are generally insufficient to provide adequate coverage of the system operable space and do not capture rare unstable events. This work presents a tool for the synthetic generation of power system operating scenarios that jointly considers optimal power flow feasibility and small-signal stability. The proposed approach enables extensive exploration of the operable space, identifies multi-class boundaries between stable, unstable, and infeasible operating points, and guides the exploration towards these regions. Different sampling strategies can be employed either to maximize sample variance or to reproduce historical demand and generation patterns. The framework also supports the exploration of different network topologies, including randomly perturbed network impedances and N-1 line outages. In addition, the tool provides exploration capabilities particularly relevant to systems with high penetration of inverter-based resources (IBRs), enabling the systematic assessment of stability under different synchronous generator/IBR power-injection mixes, grid-following/grid-forming control configurations, and controller parameter settings. Finally, the proposed workflow facilitates parallel execution, enabling high-performance computing to improve the computational scalability of large-scale data generation campaigns.

Abstract

The emergence of graph neural networks (GNNs) and foundation models for power-system analysis has created a critical need for rigorous, reproducible evaluation methodologies. While AI surrogates for Power Flow (PF) and Optimal Power Flow (OPF) report significant speedups over numerical solvers, direct comparison across studies remains difficult due to differences in datasets, operating assumptions, and metrics. Furthermore, assessing readiness for practical deployment requires benchmarking these models against the established approximations and optimization methods used in industrial practice. This work presents PF and OPF surrogates benchmarking framework that emphasizes comparability across architectures, operating conditions, and grid topologies. The framework utilizes standardized datasets covering synthetic and benchmark transmission networks up to 2,000 buses, including contingency scenarios, feasible and near-infeasible operating regimes. We evaluate representative architecture families—including message-passing GNNs, heterogeneous graph transformers, iterative neural solvers, and physics-informed decoders—against classical numerical baselines through a common evaluation harness. A key contribution is an architecture-agnostic scoring pipeline that maps diverse model outputs to a standardized physical state representation to compute physics-based residuals, constraint violations, generalization metrics, and inference latency. This framework provides a reproducible foundation for transparent comparison and establishes a necessary step toward benchmark standards for AI-enabled power-system analysis.

Abstract

Learning-based surrogates for AC power-grid analysis are typically trained for one network and one task, which limits their reuse as system models evolve. This poster investigates transfer learning across networks for three coupled tasks, power flow, optimal power flow, and state estimation, using a single heterogeneous graph neural network. Buses are represented as typed nodes (PQ, PV, slack), branch admittances as edge attributes, and a one-hot task indicator within the node features selects the formulation, so that one model predicts voltage magnitude and angle, generator active and reactive power, and active and reactive demand. The backbone is a stack of residual gated graph convolutions with type-specific message passing and a shared readout. Training data are generated with gridfm-datakit for three PGLib benchmark networks: case118 and case500 constitute the pretraining pool, while case6470 is held out entirely as the transfer target. Each operating scenario is expanded into the three task formulations and verified against the AC power-balance equations prior to training. We compare zero-shot evaluation, full fine-tuning, and training on the target network alone under matched sample budgets. Fine-tuning attains a coefficient of determination of 0.91 for voltage magnitude and 0.85 for voltage angle on the held-out network.