Enhancing Weather-Related Outage Prediction and Precursor Discovery Through Attention-Based Multi-Level Modeling.

Electric grid continually monitors spatiotemporal data from sparse service areas. As power systems grow and get more complex, and with the deployment of more sensors and data collection capabilities, monitoring and analyzing data streams for outage prediction will get more complicated. In addition, the burden on human operators to analyze such data is getting challenging. Furthermore, climate change introduces new challenges to power grid reliability and makes the human grid operators' task more critical. To address some of these challenges, this research proposes a novel model to jointly predict power grid outages and discover precursors from spatiotemporal data using multi-level data. The new method utilizes multi-task learning (MTL) and multi-instance learning (MIL) to jointly predict outages and learn event precursors. This is achieved by introducing distance-aware self-attention to capture relationships between locations and improve event detection and precursor discovery while utilizing multi-level data (local weather data, global demand, and forecast data) in a sparse setting. Experiments are conducted using five years of data collected in the U.S. Pacific Northwest. The proposed methodology achieves an Area Under the Precision-Recall Curve (AU-PRC) of 0.97 using 12 hours of data before the event. Experiments showed that the proposed model could predict events several hours ahead with high accuracy, where such early predictions allow grid operators to deploy outage mitigation plans. In addition, the new framework effectively discovers spatiotemporal precursors for power outages. Grid operators can use such event precursors to help mitigate outages and improve grid reliability.