Beyond the Blackbox: An Evidence-Based Framework for Power Outage Prediction and Cross-Continent Transfer Learning Practice

An interpretable machine learning system based on XGBoost that applies real U.S. power outage data (EAGLE-I dataset) to weather-induced outage prediction in India's UP/NCR region via cross-continent transfer learning, and achieves precise risk assessment by integrating infrastructure vulnerability scores. The project addresses the lack of outage data in India and provides a scientific basis for power system management.

In power system management, outage prediction has long relied on simple weather threshold rules (e.g., predicting outages when wind speed exceeds 60 km/h), but these fail to capture complex non-linear relationships (such as the impact of sustained high humidity and moderate high temperatures on transformer lifespan). The Beyond-the-Blackbox project, developed by Amisha Srivastava's team, aims to build an evidence-based interpretable machine learning framework. Unlike traditional complex neural networks, the project establishes a decision classification system based on temporal resolution and data availability, which is grounded in a systematic review of 113 case studies from 41 academic papers.

There is no public outage dataset for India's UP/NCR region (cities like Lucknow and Noida), making direct training of a localized model impossible. The team's key insight: The physical laws of grid failures are universal (e.g., transformer overheating due to thermal stress, transmission line breakage from strong winds). Thus, they adopted a cross-continent transfer learning strategy—training the model with real U.S. outage data and applying it to prediction in India.

First Stage: U.S. Data Preparation

Obtain 2023 county-level outage events (15-minute resolution, 26 million rows) from the U.S. Department of Energy's EAGLE-I dataset, and get matching hourly weather data via the Open-Meteo Archive API. Preprocessing steps include: filtering 6 U.S. states with similar climates (e.g., Texas), acquiring weather data for 20 cities, fusing data using Haversine distance matching, and engineering 13 initial features (v1).

Second Stage: Model Training and Optimization

XGBoost is used (efficient for tabular data, built-in feature importance, supports cost-sensitive learning). Training uses a cost-sensitive strategy (scale_pos_weight=6.37) with two iterations:

• v1 model: 13 features (heat index, season markers, etc.)
• v2 model: 13 additional features (gusts, rolling temperature, etc.), validated effective by MRMR

v2 model performance: accuracy 74.4%, recall 51.6%, precision 27.0%, F1=0.354. Model tuning prioritizes high recall (reducing missed alerts) at the cost of low precision (more false alerts).

Two adaptations are needed for transfer to India's UP/NCR:

1. Season definition adjustment: India's summer is April-June (not June-August in the Northern Hemisphere), affecting the calculation of is_summer and month features.
2. Infrastructure vulnerability score: Referencing Wang et al. (2024), calculate city vulnerability multipliers based on official DISCOM distribution loss data from UPERC/PFC (2023-24 fiscal year):

City	DISCOM	Rating	Vulnerability Score	Impact on 45% Original Risk
Noida	PVVNL	A+	0.93	→41.9%
Ghaziabad	PVVNL	A+	1.00	→45.0%
Meerut	PVVNL	A+	1.07	→48.2%
Lucknow	MVVNL	B-	1.13	→50.9%
Agra	DVVNL	B-	1.27	→57.2%
Firozabad	DVVNL	B-	1.40	→63.0%

Under the same weather conditions, cities with poorer infrastructure have higher outage rates.

A four-level risk classification system is established:

• 🟢 Low risk (<30%): Grid safe
• 🟡 Medium risk (30-50%): Increase vigilance
• 🟠 High risk (50-70%): Prepare emergency plans
• 🔴 Extreme risk (≥70%): Activate emergency response

Key predictive features: temp_x_humidity (combined heat and humidity stress), is_summer (highest risk season), month (seasonal pattern), surface_pressure (low pressure indicates storms), is_monsoon (monsoon period marker).

Limitations

1. Cross-continent transfer gap: The U.S. model learns high-temperature-dominated failures, while India's failure mechanisms are different (rainfall, overloaded transformers, poor maintenance).
2. Pure weather features: Lack of utility data (equipment age, maintenance records, etc.) limits accuracy.
3. Precision trade-off: High recall leads to alert fatigue.

Future Directions

• Integrate LSTM time-series models for sequence prediction
• Use graph neural networks to model spatial cascading effects of substations
• Supplement utility-specific data to improve accuracy

The project provides lessons for critical infrastructure prediction:

1. Data scarcity can be mitigated via transfer learning + domain knowledge
2. Interpretability (e.g., XGBoost feature importance) helps operators understand prediction basis
3. Localization adjustments are more effective than global models (e.g., vulnerability scores)
4. Cost-sensitive design must match business scenarios (prioritize recall for safety)

For developing countries: Use open data to train basic models, combine with local knowledge for refined adjustments, and implement practical prediction systems.