Power outage prediction using data streams: An adaptive ensemble learning approach with a feature- and...

Area of Science:

• Environmental Science
• Computer Science
• Electrical Engineering

Background:

• Weather events like windstorms and heatwaves significantly disrupt power systems, causing outages and inconveniences.
• Accurate prediction of customer power outages is crucial for efficient utility restoration efforts.
• Current models struggle with data streams and model uncertainty from diverse weather event data.

Purpose of the Study:

• To develop an adaptive, all-weather power outage prediction model capable of handling data streams.
• To address the limitations of existing models in managing diverse weather data and associated uncertainties.

Main Methods:

• Proposed an adaptive ensemble learning algorithm designed for data streams.
• Implemented a feature- and performance-based weighting mechanism to combine base learner outputs.
• Utilized a large, real-world dataset of daily customer interruptions for model development and validation.

Main Results:

• The proposed algorithm demonstrated more accurate probabilistic predictions compared to benchmark approaches.
• Achieved a reduction in probabilistic prediction error ranging from 4% to 22%, with an average reduction of 8%.
• Observed enhanced performance improvements when simpler models were used as base learners.

Conclusions:

• The developed adaptive ensemble learning algorithm effectively predicts power outages from data streams.
• The model offers significant improvements in probabilistic and point outage predictions, aiding utility response.
• This approach represents a novel solution for all-weather outage prediction in power systems.