arxiv: 2604.04916 · v1 · submitted 2026-04-06 · 💻 cs.LG

Recognition: no theorem link

Empowering Power Outage Prediction with Spatially Aware Hybrid Graph Neural Networks and Contrastive Learning

Xuyang Shen , Zijie Pan , Diego Cerrai , Xinxuan Zhang , Christopher Colorio , Emmanouil N. Anagnostou , Dongjin Song

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:29 UTC · model grok-4.3

classification 💻 cs.LG

keywords power outage predictiongraph neural networkscontrastive learningextreme weatherspatial awarenessutility networksoutage modelingmachine learning

0 comments

The pith

SA-HGNN with contrastive learning achieves state-of-the-art accuracy in predicting power outages from extreme weather.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops Spatially Aware Hybrid Graph Neural Networks combined with contrastive learning to improve forecasts of power outages triggered by severe storms, hurricanes, and snow events. It encodes spatial relationships among fixed landscape features and changing weather conditions while using contrastive methods to create location-specific embeddings that balance different event types. This matters because climate change is increasing such events, and better advance predictions let utilities and communities prepare to limit damage to infrastructure, economies, and daily life. Tests across four northeastern utility territories show the approach outperforms prior methods in the outage prediction system.

Core claim

We develop Spatially Aware Hybrid Graph Neural Networks (SA-HGNN) with contrastive learning to enhance the OPM predictions for extreme weather-induced power outages. We first encode spatial relationships of both static features (e.g., land cover, infrastructure) and event-specific dynamic features (e.g., wind speed, precipitation) via SA-HGNN. Next, we leverage contrastive learning to handle the imbalance problem associated with different types of extreme weather events and generate location-specific embeddings by minimizing intra-event distances between similar locations while maximizing inter-event distances across all locations. Thorough empirical studies in four utility service areas, i.

What carries the argument

Spatially Aware Hybrid Graph Neural Networks (SA-HGNN) paired with contrastive learning to encode spatial relationships of static and dynamic features and produce location-specific embeddings.

If this is right

Pre-emptive forecasts become available for electric distribution networks ahead of weather events.
Imbalance across storm types is handled more effectively through location-specific embeddings.
State-of-the-art prediction performance holds in Connecticut, Western Massachusetts, Eastern Massachusetts, and New Hampshire.
Disruptions to industrial operations, communities, and critical infrastructure can be reduced through earlier preparation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar spatial graph and contrastive techniques could extend to predicting weather impacts on transportation or water infrastructure.
The approach might adapt to other imbalanced forecasting tasks in environmental or climate modeling.
Validation on datasets from regions outside the northeastern U.S. would test broader geographic applicability.
Explicit spatial awareness could be added to other existing outage prediction systems to address their current limitations.

Load-bearing premise

The spatial encodings and contrastive embeddings learned from training data will generalize to new extreme weather events and territories without overfitting to the specific distributions or feature choices in the four studied areas.

What would settle it

A direct comparison showing that SA-HGNN underperforms existing models when applied to outage data from a fifth utility territory or an extreme weather event type absent from the original training sets.

Figures

Figures reproduced from arXiv: 2604.04916 by Christopher Colorio, Diego Cerrai, Dongjin Song, Emmanouil N. Anagnostou, Xinxuan Zhang, Xuyang Shen, Zijie Pan.

**Figure 2.** Figure 2: Actual outages vs. predicted outages comparison of four models on Connecticut extreme weather [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison of learned location embeddings across different methods on the Connecticut dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

**Figure 4.** Figure 4: SA-HGNN CL representation comparison w/o CL: SA-HGNN without contrastive learning, where the SA-HGNN is trained to generate location embeddings without explicitly grouping similar location pairs or enforcing discrimination between different types of locations. The results show that removing contrastive learning degrades model performance, underscoring its role in enhancing outage prediction. In addition… view at source ↗

**Figure 5.** Figure 5: Outage prediction distribution of Hurricane Irene (August 28, 2011) in Connecticut. The ground [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗

**Figure 6.** Figure 6: Outage prediction distribution of Storm (October 29, 2017) in East Massachusetts. The ground [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 7.** Figure 7: Hyperparameter sensitivity analysis. We report MAPE, AE q25, and APE q25 under different [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗

read the original abstract

Extreme weather events, such as severe storms, hurricanes, snowstorms, and ice storms, which are exacerbated by climate change, frequently cause widespread power outages. These outages halt industrial operations, impact communities, damage critical infrastructure, profoundly disrupt economies, and have far-reaching effects across various sectors. To mitigate these effects, the University of Connecticut and Eversource Energy Center have developed an outage prediction modeling (OPM) system to provide pre-emptive forecasts for electric distribution networks before such weather events occur. However, existing predictive models in the system do not incorporate the spatial effect of extreme weather events. To this end, we develop Spatially Aware Hybrid Graph Neural Networks (SA-HGNN) with contrastive learning to enhance the OPM predictions for extreme weather-induced power outages. Specifically, we first encode spatial relationships of both static features (e.g., land cover, infrastructure) and event-specific dynamic features (e.g., wind speed, precipitation) via Spatially Aware Hybrid Graph Neural Networks (SA-HGNN). Next, we leverage contrastive learning to handle the imbalance problem associated with different types of extreme weather events and generate location-specific embeddings by minimizing intra-event distances between similar locations while maximizing inter-event distances across all locations. Thorough empirical studies in four utility service territories, i.e., Connecticut, Western Massachusetts, Eastern Massachusetts, and New Hampshire, demonstrate that SA-HGNN can achieve state-of-the-art performance for power outage prediction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

SA-HGNN plus contrastive learning gives measurable gains on the four New England territories but the generalization claim rests on in-sample results without clear hold-outs for rare events.

read the letter

The paper takes hybrid graph networks, adds explicit spatial encoding for both static infrastructure and dynamic weather fields, and layers on contrastive learning to pull similar locations closer while pushing different event types apart. That setup is then plugged into an existing outage prediction pipeline at Eversource. The empirical work covers four adjacent service areas and reports better numbers than the prior models they had in production. The contrastive objective is a reasonable way to handle the fact that hurricanes, snowstorms, and ice events do not occur at equal rates, and the spatial hybrid layer lets the model use land-cover and grid topology without flattening everything into a single feature vector. Those pieces are straightforward and the authors execute them cleanly on real utility data. The main limitation is that all four territories share similar climate, vegetation, and grid characteristics, and the evaluation does not appear to include a temporal hold-out of the most extreme storms or a leave-one-territory-out test. Without those checks it is hard to know whether the learned embeddings capture transferable structure or simply fit the particular event frequencies and spatial autocorrelations in the training windows. The abstract also gives no error bars, no details on the exact baselines, and no discussion of how spatial leakage was avoided in the splits. This work is aimed at applied researchers who already run outage models for utilities or who want a concrete geospatial ML example for infrastructure resilience. A reader who needs a deployable improvement for similar mid-latitude service territories will find usable ideas here. I would send it to peer review; the core architecture is sensible and the results on the given data are worth a closer look, even if the authors will need to add stronger out-of-distribution tests before the SOTA claim can be taken as settled.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces Spatially Aware Hybrid Graph Neural Networks (SA-HGNN) combined with contrastive learning to predict power outages from extreme weather events. Static features (land cover, infrastructure) and dynamic features (wind speed, precipitation) are encoded via the hybrid GNN to capture spatial relationships; contrastive learning then produces location-specific embeddings by minimizing intra-event distances and maximizing inter-event distances to address class imbalance. The central claim is that thorough empirical studies across four New England utility territories (Connecticut, Western Massachusetts, Eastern Massachusetts, New Hampshire) establish state-of-the-art performance for outage prediction.

Significance. If the empirical claims are substantiated with complete experimental protocols, the work could meaningfully advance operational outage prediction models by explicitly incorporating spatial structure and event-type contrastive objectives, offering utilities better pre-event forecasts. The hybrid GNN plus contrastive formulation is a natural extension of existing graph-based spatial modeling and self-supervised techniques to the power-outage domain.

major comments (2)

[Abstract] Abstract: the claim that 'Thorough empirical studies in four utility service territories... demonstrate that SA-HGNN can achieve state-of-the-art performance' is unsupported because the abstract (and, per the provided description, the methods) supplies no information on the baselines, metrics, cross-validation procedure, error bars, or controls for spatial autocorrelation.
[Empirical Studies] Empirical evaluation: no temporal hold-out for rare extreme events, no leave-one-territory-out evaluation, and no external-territory test are reported. Without these, the reported superiority cannot be distinguished from overfitting to the particular event frequencies and spatial autocorrelations present in the four training territories.

minor comments (1)

[Abstract] The abstract would be clearer if it briefly named the performance metrics and the magnitude of improvement over the strongest baseline.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on clarifying our empirical claims and evaluation procedures. We address each major comment below and will revise the manuscript to improve transparency and robustness where feasible.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that 'Thorough empirical studies in four utility service territories... demonstrate that SA-HGNN can achieve state-of-the-art performance' is unsupported because the abstract (and, per the provided description, the methods) supplies no information on the baselines, metrics, cross-validation procedure, error bars, or controls for spatial autocorrelation.

Authors: We agree that the abstract is too concise to fully support the SOTA claim on its own. The full manuscript (Section 4) specifies the baselines (logistic regression, random forest, XGBoost, standard GNN variants), metrics (precision, recall, F1, AUC-ROC), temporal cross-validation respecting event chronology, and reporting of means with standard deviations. Spatial autocorrelation is addressed by construction through the hybrid graph that encodes location dependencies. We will revise the abstract to include a brief statement on the evaluation protocol and statistical reporting to make the claim self-contained. revision: yes
Referee: [Empirical Studies] Empirical evaluation: no temporal hold-out for rare extreme events, no leave-one-territory-out evaluation, and no external-territory test are reported. Without these, the reported superiority cannot be distinguished from overfitting to the particular event frequencies and spatial autocorrelations present in the four training territories.

Authors: We acknowledge the value of these additional checks for rare-event generalization. The existing protocol already uses a strict temporal hold-out (earlier events for training, later events including extremes for testing) within each territory. We did not originally include leave-one-territory-out or external-territory tests because the study is confined to the four territories with harmonized data. In revision we will add leave-one-territory-out results and expand the discussion to explicitly address potential overfitting to local event frequencies and spatial structure, while noting the data limitation that precludes external-territory testing. revision: partial

Circularity Check

0 steps flagged

No circularity in model derivation or performance claims

full rationale

The paper presents SA-HGNN as a hybrid GNN architecture augmented with contrastive learning for spatial feature encoding and imbalance handling in outage prediction. All performance claims rest on standard empirical evaluation across four territories using supervised training and testing splits. No equations, parameters, or results are defined in terms of the target metrics themselves, no predictions reduce to fitted inputs by construction, and no load-bearing steps rely on self-citations or uniqueness theorems imported from the authors' prior work. The derivation chain is a conventional ML pipeline whose outputs are independently falsifiable via hold-out data.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on standard assumptions of graph neural network expressivity for spatial data and the effectiveness of contrastive learning for handling event-type imbalance; no explicit free parameters, axioms, or invented entities are stated in the abstract.

pith-pipeline@v0.9.0 · 5590 in / 1094 out tokens · 56574 ms · 2026-05-10T19:29:08.993228+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

30 extracted references · 27 canonical work pages · 1 internal anchor

[1]

Celestial Mechan- ics and Dynamical Astronomy83, 155–169 (2002) https://doi.org/10.1023/A: 1020143116091

ISSN 0885-6125. doi: 10.1023/A: 1010933404324. URLhttps://doi.org/10.1023/A:1010933404324. Richard J. Campbell. Weather-related power outages and electric system resiliency.Technical Report, Congressional Research Service,

work page doi:10.1023/a:
[2]

Diego Cerrai, Md Abul Ehsan Bhuiyan, Xinxuan Zhang, Jaemo Yang, Maria Frediani, and Emmanouil Anagnostou

URLhttps://api.semanticscholar.org/CorpusID:108180774. Diego Cerrai, Md Abul Ehsan Bhuiyan, Xinxuan Zhang, Jaemo Yang, Maria Frediani, and Emmanouil Anagnostou. Predicting storm outages through new representations of weather and vegetation.IEEE Access, PP:1–1, 03 2019a. doi: 10.1109/ACCESS.2019.2902558. Diego Cerrai, Marika Koukoula, Peter Watson, and Emm...

work page doi:10.1109/access.2019.2902558 2019
[3]

In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

Association for Computing Machinery. ISBN 9781450342322. doi: 10.1145/2939672. 2939785. URLhttps://doi.org/10.1145/2939672.2939785. Yuzhou Chen, Ignacio Segovia, and Yulia R Gel. Z-gcnets: Time zigzags at graph convolutional networks for time series forecasting. InInternational Conference on Machine Learning, pp. 1684–1694. PMLR,

work page doi:10.1145/2939672
[4]

A systematic literature review of spatio-temporal graph neural network models for time series forecasting and classification.arXiv preprint arXiv:2410.22377,

Flavio Corradini, Flavio Gerosa, Marco Gori, Carlo Lucheroni, Marco Piangerelli, and Martina Zannotti. A systematic literature review of spatio-temporal graph neural network models for time series forecasting and classification.arXiv preprint arXiv:2410.22377,

work page arXiv
[5]

Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst

doi: 10.1016/ j.crm.2019.100193. Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering.Advances in neural information processing systems, 29,

work page arXiv 2019
[6]

Nina Flores, Heather McBrien, Vivian Do, Mathew Kiang, Jeffrey Schlegelmilch, and Joan Casey

doi: 10.1289/EHP2154. Nina Flores, Heather McBrien, Vivian Do, Mathew Kiang, Jeffrey Schlegelmilch, and Joan Casey. The 2021 texas power crisis: distribution, duration, and disparities.Journal of Exposure Science & Environmental Epidemiology, 33:1–11, 08

work page doi:10.1289/ehp2154 2021
[7]

Jasmine Garland, Kyri Baker, and Ben Livneh

doi: 10.1038/s41370-022-00462-5. Jasmine Garland, Kyri Baker, and Ben Livneh. Weather-induced power outage prediction: A comparison of machine learning models. In2023 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm), pp. 1–6. IEEE,

work page doi:10.1038/s41370-022-00462-5
[8]

doi: 10.1201/9781351174664-353

ISBN 9781351174664. doi: 10.1201/9781351174664-353. Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pp. 1024–1034,

work page doi:10.1201/9781351174664-353
[9]

Kaveh Hassani and Amir Hosein Khasahmadi

doi: 10.1109/ TGRS.2021.3100847. Kaveh Hassani and Amir Hosein Khasahmadi. Contrastive multi-view representation learning on graphs. In Hal Daumé III and Aarti Singh (eds.),Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pp. 4116–4126. PMLR, 13–18 Jul

work page arXiv 2021
[10]

URL https://ideas.repec.org/a/wly/riskan/v37y2017i3p441-458.html

doi: 10.1111/risa.12652. URL https://ideas.repec.org/a/wly/riskan/v37y2017i3p441-458.html. Noah Hollmann, Samuel Müller, Lennart Purucker, Arjun Krishnakumar, Max Körfer, Shi Bin Hoo, Robin Tibor Schirrmeister, and Frank Hutter. Accurate predictions on small data with a tabular foun- dation model.Nature, 01

work page doi:10.1111/risa.12652
[11]

Accurate predictions on small data with a tabular foundation model.Nature, 637(8045):319–326, 2025

doi: 10.1038/s41586-024-08328-6. URLhttps://www.nature.com/ articles/s41586-024-08328-6. 16 Published in Transactions on Machine Learning Research (04/2026) Haojun Jiang, Jiawei Sun, Jie Li, and Chentao Wu. Localgcl: Local-aware contrastive learning for graphs. arXiv preprint arXiv:2402.17345,

work page doi:10.1038/s41586-024-08328-6 2026
[12]

Thomas N Kipf and Max Welling

doi: 10.1145/3711896.3736567. Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907,

work page doi:10.1145/3711896.3736567
[13]

Rethinking domain generaliza- tion: Discriminability and generalizability.arXiv preprint arXiv:2309.16483,

Shaocong Long, Qianyu Zhou, Chenhao Ying, Lizhuang Ma, and Yuan Luo. Rethinking domain generaliza- tion: Discriminability and generalizability.arXiv preprint arXiv:2309.16483,

work page arXiv
[14]

doi: 10.1111/risa. 12131. URLhttps://ideas.repec.org/a/wly/riskan/v34y2014i6p1069-1078.html. Damian Owerko, Fernando Gama, and Alejandro Ribeiro. Predicting power outages using graph neural networks. In2018 ieee global conference on signal and information processing (globalsip), pp. 743–747. IEEE,

work page doi:10.1111/risa
[15]

Structural knowledge informed continual multivariate time series forecasting.arXiv preprint arXiv:2402.12722,

Zijie Pan, Yushan Jiang, Dongjin Song, Sahil Garg, Kashif Rasul, Anderson Schneider, and Yuriy Nevmy- vaka. Structural knowledge informed continual multivariate time series forecasting.arXiv preprint arXiv:2402.12722,

work page arXiv
[16]

doi: 10.1175/BAMS-D-15-00308

work page doi:10.1175/bams-d-15-00308
[17]

doi: 10.1109/CVPR.2019. 01230. URLhttps://doi.ieeecomputersociety.org/10.1109/CVPR.2019.01230. W.C. Skamarock, J. Klemp, Jimy Dudhia, D.O. Gill, Dale Barker, Wei Wang, and J.G. Powers. A description of the advanced research wrf version

work page doi:10.1109/cvpr.2019 2019
[18]

A.B. Smith. U.s. billion-dollar weather and climate disasters (2025).NOAA National Centers for Environ- mental Information (NCEI),

2025
[19]

17 Published in Transactions on Machine Learning Research (04/2026) Nengli Sun, Zeming Zhou, Qian Li, and Jinrui Jing

URLhttps://openreview.net/forum?id=r1lfF2NYvH. 17 Published in Transactions on Machine Learning Research (04/2026) Nengli Sun, Zeming Zhou, Qian Li, and Jinrui Jing. Three-dimensional gridded radar echo extrapolation for convective storm nowcasting based on 3d-convlstm model.Remote Sensing, 14:4256, 08

2026
[20]

Shanshan Tang, Bo Li, and Haijun Yu

doi: 10.3390/rs14174256. Shanshan Tang, Bo Li, and Haijun Yu. Chebnet: efficient and stable constructions of deep neural networks with rectified power units via chebyshev approximation.Communications in Mathematics and Statistics, pp. 1–27,

work page doi:10.3390/rs14174256
[21]

Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio

doi: 10.1109/ ACCESS.2024.3446311. Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks.6th International Conference on Learning Representations,

work page arXiv 2024
[22]

URLhttps://ideas.repec.org/a/spr/nathaz/v79y2015i2p1359-1384.html

doi: 10.1007/ s11069-015-1908-2. URLhttps://ideas.repec.org/a/spr/nathaz/v79y2015i2p1359-1384.html. Peter Watson, Aaron Spaulding, Marika Koukoula, and Emmanouil Anagnostou. Improved quantitative prediction of power outages caused by extreme weather events.Weather and Climate Extremes, 37:100487, 07

1908
[23]

Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang

doi: 10.1016/j.wace.2022.100487. Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial- temporal graph modeling.International Joint Conferences on Artificial Intelligence Organization, pp. 1907–1913,

work page doi:10.1016/j.wace.2022.100487 2022
[24]

Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence,

URLhttps://doi.org/10.24963/ijcai.2019/264. Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, Xiaojun Chang, and Chengqi Zhang. Connecting the dots: Multivariate time series forecasting with graph neural networks. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’20, pp. 753–763, New York, NY, USA,

work page doi:10.24963/ijcai.2019/264 2019
[25]

LayoutLM: Pre-training of Text and Layout for Document Image Understanding

Association for Computing Machinery. ISBN 9781450379984. doi: 10.1145/3394486. 3403118. URLhttps://doi.org/10.1145/3394486.3403118. Yaochen Xie, Zhao Xu, Jingtun Zhang, Zhengyang Wang, and Shuiwang Ji. Self-supervised learning of graph neural networks: A unified review.IEEE transactions on pattern analysis and machine intelligence, 45 (2):2412–2429,

work page doi:10.1145/3394486
[26]

How Powerful are Graph Neural Networks?

Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? CoRR, abs/1810.00826,

work page internal anchor Pith review arXiv
[27]

Feifei Yang, Diego Cerrai, and Emmanouil N Anagnostou

doi: 10.1109/ ACCESS.2020.2983159. Feifei Yang, Diego Cerrai, and Emmanouil N Anagnostou. The effect of lead-time weather forecast uncer- tainty on outage prediction modeling.Forecasting, 3(3):501–516,

work page arXiv 2020
[28]

Regexplainer: Generating expla- nations for graph neural networks in regression task.arXiv preprint arXiv:2307.07840,

Jiaxing Zhang, Zhuomin Chen, Hao Mei, Dongsheng Luo, and Hua Wei. Regexplainer: Generating expla- nations for graph neural networks in regression task.arXiv preprint arXiv:2307.07840,

work page arXiv
[29]

Continual learning on graphs: Challenges, solutions, and opportunities.arXiv preprint arXiv:2402.11565, 2024

18 Published in Transactions on Machine Learning Research (04/2026) Michael Zhang, Nimit S. Sohoni, Hongyang R. Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. InInternational Conference on Machine Learning (ICML), 2022a. Rui Zhang, Yunxing Zhang, Chengjun Lu, and Xuelo...

work page arXiv 2026
[30]

doi: 10.1109/ TPAMI.2022.3195549. 19 Published in Transactions on Machine Learning Research (04/2026) A Selected Feature Information This section provides a comprehensive list of all the selected features along with their explanations. Each feature is described in detail to clarify its significance and relevance to the analysis. These features encompass v...

work page arXiv 2022