Uncovering Contributing Factors to Interruptions in the Power Grid: An Arctic Case

Odin Foldvik Eikeland, Filippo Maria Bianchi, Inga Setså Holmstrand, Sigurd Bakkejord, Sérgio F. Santos, Matteo Chiesa

TL;DR

This study applies machine learning to a real-world Arctic power grid to predict short-term failures that disrupt critical local industries. By combining high-resolution topographical data with meteorological and industrial load patterns, the researchers achieved a 57% success rate in predicting failures, identifying wind speed and industrial activity as the primary triggers.

What Problem Does This Paper Address?

In the Arctic region of North Norway, power reliability isn't just a matter of convenience—it’s the backbone of the local economy, specifically the fish processing industry. The SVAN22LY1 grid is a "radial" system, meaning it lacks redundant paths; if one line fails, the customers at the end of the line lose power completely. Even a momentary voltage dip can shut down automated production lines for an hour, leading to significant economic losses.

While most research focuses on weather-related outages using benchmark datasets, this paper tackles a messy, real-world challenge: identifying why failures happen in a specific, harsh environment where both extreme weather and heavy industrial loads collide. The goal is to move from reactive repairs to proactive mitigation by understanding the "where" and "why" of grid instability.

Methods

The researchers developed a multi-step pipeline to analyze the grid:

  1. Topographical Mapping: They mapped 978 utility poles to a 1-meter resolution Digital Surface Model (DSM) to identify "Exposed Areas" (EAs)—regions at high altitudes with long spans between poles where wind is most likely to cause line collisions.
  2. Statistical Feature Selection: They used Kernel Density Estimation (KDE) and Kullback-Leibler (KL) divergence to determine which variables (weather vs. load) actually helped distinguish between normal operations and failure states.
  3. Machine Learning: Because failures are rare (only 44 incidents out of ~8,000 hours of data), the authors faced a massive class imbalance. They tested several models:
  4. One-Class Classifiers: (Isolation Forest, etc.) to detect anomalies.
  5. Standard SVM: Which failed completely by ignoring the rare failure events.
  6. Weighted SVM: A modified Support Vector Machine that penalizes missing a failure (False Negative) 10 times more than a false alarm (False Positive).
  7. SMOTE-SVM: Using synthetic oversampling to balance the dataset.

Datasets

The study utilized a diverse set of data from 2020:

  • Grid Topology: Locations and altitudes of 978 utility poles provided by Arva Power Company.
  • FASIT Reports: The official Norwegian logging system for power interruptions.
  • AROME-Arctic Weather Model: Hourly data (wind, temp, pressure, etc.) at a 2.5 km spatial resolution.
  • Energy Data: Hourly consumption from local industry, hydropower production levels, and data from "Transformer South," which monitors broader grid strain.
  • External Industry Data: High-resolution (1-minute) consumption data from a massive industrial customer 30km away that puts heavy strain on the regional grid.

Key Insights and Ideas

  • The "Double Whammy" of Wind and Load: The "aha moment" was confirming that wind speed and industrial activity are nearly equal contributors to grid failure. High wind speeds (above 25 m/s) at high altitudes are the primary physical trigger, but the fluctuation in industrial power demand (the difference between min and max power within an hour) significantly lowers the grid's resilience.
  • Topography as a Filter: By focusing on altitude and pole-to-pole distance, the researchers could narrow down potential failure points to three specific "Exposed Areas," providing a roadmap for physical grid reinforcement.
  • Weighting vs. Oversampling: In this specific case, simply weighting the SVM model was more effective than using SMOTE (Synthetic Minority Over-sampling Technique). This suggests that for grid failures, the "cost" of a missed detection is the most critical parameter to tune in the algorithm.

Key Results

  • Predictive Power: The Weighted SVM was the top performer, achieving an F1-score of 0.53 and a sensitivity of 57%. This means it correctly identified over half of the failures in the test set while only producing 4 false positives.
  • Top Predictors: Wind speed had the highest KL divergence (3.87), followed closely by local industry energy consumption (3.83).
  • Physical Confirmation: Individual investigation of the 7 failures in the test set confirmed that high wind (over 25 m/s) was present during most incidents, validating the model's focus.

Limitations and Future Work

  • Temporal Resolution: The current model uses hourly data. However, power failures often happen due to "transient" events lasting seconds. The authors note that moving to minute-by-minute data could significantly improve accuracy but requires more computational power and data access.
  • Localization: While the study identifies regions at risk, it does not yet pinpoint the exact utility pole or component that triggered a specific failure.
  • Unexplained Failures: Two out of the seven test failures remained "unknown" in cause, suggesting there are still hidden variables (perhaps equipment age or specific icing events) not captured in the current dataset. Future work will focus on integrating higher-resolution data to localize these "weather-related causing factors" more precisely.

Summary updated

Abstract

Electric failures are a problem for customers and grid operators. Identifying causes and localizing the source of failures in the grid is critical. Here, we focus on a specific power grid in the Arctic region of Northern Norway. First, we collected data pertaining to the grid topology, the topography of the area, the historical meteorological data, and the historical energy consumption/production data. Then, we exploited statistical and machine-learning techniques to predict the occurrence of failures. The classification models achieve good performance, meaning that there is a significant relationship between the collected variables and fault occurrence. Thus, we interpreted the variables that mostly explain the classification results to be the main driving factors of power interruption. Wind speed of gust and local industry activity are found to be the main controlling parameters in explaining the power failure occurrences. The result could provide important information to the distribution system operator for implementing strategies to prevent and mitigate incoming failures.
BibTeX
@inproceedings{eikeland2022uncovering,
  author = {Odin Foldvik Eikeland and Filippo Maria Bianchi and Inga Setså Holmstrand and Sigurd Bakkejord and Sérgio F. Santos and Matteo Chiesa},
  title = {{Uncovering Contributing Factors to Interruptions in the Power Grid: An Arctic Case}},
  booktitle = {Energies},
  year = {2022},
  doi = {10.3390/en15010305},
}

Loading related content...

Figure