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

Xuyang Shen [email protected], Zijie Pan, Diego Cerrai, Xinxuan Zhang, Christopher Colorio, Emmanouil N. Anagnostou, Dongjin Song [email protected] · 2026

Read Empowering Power Outage Prediction with Spatially Aware Hybrid Graph Neural Networks and Contrastive Learning by Xuyang Shen [email protected], Zijie Pan, Diego Cerrai, Xinxuan Zhang, Christopher Colorio, Emmanouil N. Anagnostou, Dongjin Song [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-04916.txt.

Introduction

Shen and six colleagues improve power-outage forecasts for severe weather, where existing models in a Connecticut utility system ignore how storms vary across space. Their spatially aware hybrid graph neural network, trained with contrastive learning, adds that geographic structure to outage prediction for distribution networks. The aim is earlier, sharper warnings before storms, hurricanes, snowstorms, and ice storms, so crews and communities can prepare. Better spatial forecasts should cut disruption to industry, infrastructure, and daily life when extreme weather hits.

Catalog record

CREATORXuyang Shen [email protected], Zijie Pan, Diego Cerrai, Xinxuan Zhang, Christopher Colorio, Emmanouil N. Anagnostou, Dongjin Song [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54747
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.04916

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Empowering Power Outage Prediction with Spatially Aware Hybrid Graph Neural Networks and Contrastive Learning Xuyang Shen [email protected], Zijie Pan, Diego Cerrai, Xinxuan Zhang, Christopher Colorio, Emmanouil N. Anagnostou, Dongjin Song [email protected] 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. 1 Introduction Power outages caused by severe weather events, such as hurricanes, snowstorms, and heavy rainfall, pose significant risks to modern society by disrupting critical infrastructure and essential services across sectors like healthcare, transportation, and finance. In the United States, weather-related power outages cost the economy an estimated $18 to $70 billion annually, with the frequency and severity of billion-dollar disasters steadily increasing over the past two decades ( 6 ; 38 ) . These outages can result in significant economic losses and, in extreme cases, loss of life ( 16 ; 14 ) . For instance, over 15 months between 2011 and 2012, three major storms in Connecticut caused extensive outages, affecting hundreds of thousands of people and resulting in considerable economic losses. Similarly, Hurricane Sandy inflicted severe damage on New York, causing prolonged service disruptions and operational challenges for utility companies ( 52 ; 42 ) . Therefore, accurate forecasting of the magnitude and spatial distribution of weather-induced power outages is essential to mitigate these impacts. Such forecasts can inform evacuation strategies, improve storm response planning, and guide investments in reinforcing and upgrading electrical infrastructure ( 1 ; 7 ; 8 ; 15 ) . Although outage prediction has received increasing attention recently, existing approaches still suffer from critical limitations. Traditional machine learning methods, such as ensemble models, have shown promising results by improving prediction accuracy and robustness ( 31 ; 45 ; 22 ; 7 ; 9 ; 42 ) . However, they treat each location independently and do not explicitly model spatial relationships between geographic regions, which are crucial for understanding the spatial effect of storm impacts. Although convolutional and recurrent neural networks, including their advanced variants ( 20 ; 40 ) , can capture spatiotemporal dependencies in grid-structured data such as radar images or weather maps, their dependency on rigid Euclidean grids limits their applicability to sensor networks or outage datasets defined on irregular geographic layouts. More recently, graph-based approaches have emerged as powerful tools for modeling non-Euclidean spatial dependencies, allowing each node to represent a spatial location. Nevertheless, existing GNN-based methods ( 26 ; 19 ; 32 ; 13 ) often consider only fixed spatial structures and lack the flexibility to model event-specific spatial dynamics. Furthermore, most prior work overlooks the inherent imbalance in outage datasets, where low-impact events dominate and high-impact events, though rare, carry greater operational significance and value. These gaps motivate the development of a more spatially adaptive and representation-discriminative approach for power outage prediction. To this end, we propose Spatially Aware Hybrid Graph Neural Networks (SA-HGNN) that leverage contrastive learning to enhance outage prediction models (OPM) for extreme weather-induced power outages. Specifically, we first construct a fixed adjacency matrix to encode the spatial relationships of static features and design a dynamic graph learning module to capture and infer complex, evolving spatial dependencies across different events. We then develop SA-HGNN to integrate spatial dependencies derived from both static and dynamic features. To address the imbalance issue associated with extreme weather events of varying severity, we incorporate a contrastive learning strategy to generate location-specific embeddings. These embeddings are obtained by minimizing intra-event distances between similar locations while maximizing inter-event distances across all locations, resulting in more discriminative representations for each location. Our main contributions include: • We introduce SA-HGNN, a novel graph-based deep learning model that can effectively integrate both static and dynamic spatial dependencies to enhance outage prediction for extreme weather events. • We develop a dynamic graph learning module that can capture and infer complex, evolving spatial relationships across different locations, addressing the limitations of existing methods that rely solely on fixed spatial structures. • To tackle the imbalance issue in outage datasets, we adopt a contrastive learning strategy that learns location-specific embeddings by minimizing intra-event distances between similar locations while maximizing inter-event distances across all locations. • Our 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. 2 Related work In recent years, machine learning methods have been increasingly employed to address the challenges of forecasting weather-related power outages ( 46 ; 17 ) . One widely adopted approach is ensemble learning. For instance, ( 31 ) utilized Random Forest (RF) to predict outages caused by tropical cyclones, while ( 45 ) employed tree-based models to forecast power outages in the New England region. He and Cerrai leveraged Bayesian additive regression trees (BART) to predict outages resulting from storm events ( 22 ; 7 ; 9 ) , whereas ( 42 ) explored RF for predicting storm-induced power outages in the New York State region. These ensemble approaches help mitigate overfitting and enhance prediction accuracy by leveraging diverse decision paths and combining the outputs of multiple decision trees. However, they do not explicitly incorporate or exploit spatial information, which is critical for outage prediction as extreme weather events ( e.g. , heavy rainfall, snowstorms) typically have strong spatial effects. Convolutional and recurrent neural networks (CNNs and RNNs), including advanced hybrid architectures like ConvLSTM, have been widely adopted to capture spatiotemporal relationships in targeted datasets, including radar images and weather sequences. These models are particularly effective at capturing both spatial and temporal dependencies, which are critical for applications such as short-term weather forecasting. For instance, ( 20 ) demonstrated the use of a U-Net model for convective precipitation prediction, while ( 40 ) applied a 3D-ConvLSTM model for storm nowcasting, showcasing the strengths of these architectures in handling complex weather data. However, CNNs are inherently designed for processing data with a regular rigid structure, such as images and time series. This makes them less suited for sensor network data used in weather and

Read the complete text