PaperlayerView exact revision on arXiv

Source-linked AI summary

EnergyEminence: Source-Aware Environmental Calibration and Evaluation in a Physics-Grounded Grid Digital Twin

Huy Trinh, Michael Mai, Yu Nong

arXiv:2609.36215v1cs.LGeess.IV

TL;DR

Power-grid digital twins must couple data-driven prediction with physically meaningful state evolution while preserving environmental-observation provenance. EnergyEminence combines graph-temporal prediction, nonlinear AC simulation, shared environmental calibration, and source-disjoint evaluation. Its testbed reveals a growing-fire shortcut failure in held-out source evaluation, supporting separate environmental alerting and cascade inference.

  • Problem

    Environmental observations such as smoke or flame do not establish electrical instability, and repeated scenarios from a small video corpus can distribute source-specific visual patterns across data partitions.

    Source: §I · p. 1

  • Method

    EnergyEminence combines graph-temporal prediction and nonlinear AC simulation on a shared grid state with source-comparable environmental evidence and source-video-disjoint scenario evaluation.

    Source: A. Shared State and Dual Inference Paths · pp. 1–2, §I · p. 1, B. Source-Comparable Environmental… · p. 2

  • Results

    The 40-scenario source-disjoint test produced 10 true positives, 8 false positives, 22 true negatives, and no false negatives, with false positives concentrated in stressed non-cascading cases from an unseen growing-fire source.

    Source: B. Held-Out Source Diagnostic · p. 3

  • Takeaways & Limitations

    The findings support keeping environmental alerts and grid-only cascade inference independently testable and bringing them together at the digital-twin decision layer.

    Source: §IV · p. 3

    The evaluation uses synthetic videos, one IEEE 118-bus-style topology, and a limited source set.

    Source: §IV · p. 3

Abstract

from arXiv · show

Power-grid digital twins must combine data-driven prediction with physically meaningful state evolution while preserving the provenance of environmental observations. This paper presents an early-stage EnergyEminence testbed that couples an IEEE 118-bus-style graph-temporal predictor, nonlinear AC cascade simulation, and operator-dashboard-like temporal replay. In addition, we introduce a shared bounded calibration that converts wildfire-detection confidence and spatial extent into source-comparable wildfire interpretable and explainable evidence. We then evaluate it with visually diverse fire and hard-negative videos. Sixteen synthetic environmental videos are curated to generate 160 source-tracked grid scenarios, and a source-video-disjoint test yields 10 true positives, 8 false positives, 22 true negatives, and no false negatives. The errors occur in stressed, non-cascading scenarios conditioned on an unseen growing-fire source. Our diagnostic then reveals environmental shortcut learning that is obscured by scenario-level random splitting. The paper therefore contributes a data-centric and inspectable evaluation workflow for multimodal grid-resilience models, together with evidence supporting separation of environmental alerting from electrical cascade inference

I. INTRODUCTION

Power-grid digital twins combine prediction with physically meaningful state evolution, but environmental observations such as smoke or flame are not equivalent to electrical instability. EnergyEminence adds shared environmental calibration and source-video-disjoint evaluation to a multimodal cascade testbed.

I. INTRODUCTION

  • Environmental observations such as smoke or flame are not equivalent to an electrically unstable grid state.Source: §I · p. 1
  • Repeatedly generating synthetic grid scenarios from a small video corpus can spread source-specific visual patterns across partitions and degrade generalization.Source: §I · p. 1
  • The paper introduces source-comparable calibration from video detections to bounded environmental evidence, evaluated with no-fire, fire, and smoke-like hard-negative sources.Source: §I · p. 1
  • A source-video-disjoint scenario protocol reveals a growing-fire shortcut failure mode hidden by scenario-level random splitting.Source: §I · p. 1

A. Shared State and Dual Inference Paths

EnergyEminence represents grid and environmental conditions on a shared time-indexed topology. It applies learned graph-temporal prediction and nonlinear AC simulation to that state through distinct inference paths.

A. Shared State and Dual Inference Paths

B. Source-Comparable Environmental Evidence

Environmental detections are converted into a shared, spatially attenuated evidence scale rather than independently normalized for each source. The evidence score uses detector confidence and detection area, with the strongest detection retained fully.

B. Source-Comparable Environmental Evidence

  • The detector evidence score uses confidence, relative bounding-box area, class weight, and an area-sensitivity parameter.A representative detector output shows retained box areas and confidence scores corresponding to the score quantities.Source: B. Source-Comparable Environmental… · p. 2
  • A shared monotonic scale preserves zero evidence and avoids forcing every source independently to a maximum of one.The strongest detection is retained fully, while secondary detections are reduced by β.Source: B. Source-Comparable Environmental… · p. 2
  • The generator spatially attenuates evidence from the sampled fire location to each grid-node position.Source: B. Source-Comparable Environmental… · p. 2

C. Synthetic Corpus and Source-Aware Split

The corpus uses screened synthetic environmental videos to generate source-tracked grid scenarios and tests transfer to unseen video sources. Its design includes smoke-like hard negatives and varied fire conditions, while the feasibility study diagnoses shortcut learning rather than causal wildfire-to-cascade prediction.

C. Synthetic Corpus and Source-Aware Split

  • Representative conditions pair industrial steam and fog hard negatives with a small vegetation fire and a night wildfire.Fog and steam test smoke-like false alarms; graded fire size and night illumination vary source and severity.Source: Figure 2 · p. 2
  • 16 of 18 synthetic clips passed screening, generating 160 scenarios from ten grid trajectories per accepted source.The six labels were no fire, hard negative, small fire, growing fire, large fire, and night fire.Source: C. Synthetic Corpus and Source-Aware… · p. 2
  • The grouped split assigned each video to one partition: 8/4/4 videos for training, validation, and testing, respectively.These partitions contained 80/40/40 scenarios and shifted evaluation toward transfer to an unseen environmental sequence.Source: C. Synthetic Corpus and Source-Aware… · p. 2
  • The feasibility experiment diagnoses shortcut learning in a coupled synthetic pipeline, not causal wildfire-to-cascade prediction.The generator intentionally correlates video evidence with grid stress.Source: C. Synthetic Corpus and Source-Aware… · pp. 2–3

A. Calibration Audit

The three-video audit compares legacy per-video normalization with a shared mapping, which avoids forcing detected fire sources to a unit maximum. The audit verifies removal of that artifact, not physical calibration or improved grid prediction.

A. Calibration Audit

  • Under the shared mapping, detected fire sources retain source-dependent maxima instead of being forced to one.The audit used α = 0.5, β = 0.25, and γ = 4.Source: A. Calibration Audit · p. 3
  • The no-fire sequence remains at zero under the shared mapping.Source: A. Calibration Audit · p. 3
  • The audit verifies removal of the forced-unit-maximum artifact, but does not establish physical fire calibration or improved grid prediction.Source: A. Calibration Audit · p. 3

B. Held-Out Source Diagnostic

The 40-scenario source-disjoint test detected all cascade cases conditioned on an unseen night-fire source, while its false positives clustered in stressed, non-cascading scenarios from an unseen growing-fire source.

B. Held-Out Source Diagnostic

IV. CONCLUSION

EnergyEminence combines graph-temporal cascade prediction, nonlinear AC simulation, environmental perception, and synchronized replay in a physics-grounded digital-twin testbed. Shared calibration and source-video-disjoint evaluation expose a growing-fire shortcut, supporting separate testing of environmental alerts and grid-only inference.

IV. CONCLUSION

  • EnergyEminence combines graph-temporal cascade prediction, nonlinear AC simulation, environmental perception, and synchronized replay in a physics-grounded digital-twin testbed.Source: §IV · p. 3
  • Shared environmental-evidence calibration improved cross-video comparability, while source-video-disjoint evaluation exposed a growing-fire shortcut hidden by scenario-level random splitting.Source: §IV · p. 3
  • The conclusion argues that environmental hazard evidence should not be treated as direct cascade ground truth; alerts and grid-only GNN/LSTM inference should remain independently testable and meet at the decision layer.Source: §IV · p. 3
  • The evaluation is limited to synthetic videos, one IEEE 118-bus-style topology, and a limited source set.The paper presents the study as a reproducible, data-driven evaluation methodology and motivates future utility-scale validation.Source: §IV · p. 3
Loading 2609.36215v1…