Geospatial mapping of distribution grid with machine learning and publicly-accessible multi-modal data.

Area of Science:

• Electrical Engineering
• Computer Science
• Geospatial Analysis

Background:

• Accurate distribution grid information is crucial for power system applications like renewable energy integration and wildfire risk assessment.
• Existing methods lack generalizability and scalability for obtaining detailed grid data.
• A need exists for a robust framework to map both overhead and underground power grids.

Purpose of the Study:

• To develop a machine learning framework for mapping distribution grids using multi-modal, publicly available data.
• To assess the framework's precision, recall, and generalizability across different geographic regions.
• To evaluate the framework's utility in estimating grid exposure to wildfires.

Main Methods:

• Utilized a machine learning approach integrating street view images, road networks, and building maps.
• Trained and benchmarked the framework against utility-owned distribution grid data in California.
• Tested the framework's generalizability by applying it to data from Sub-Saharan Africa without fine-tuning.

Main Results:

• Achieved over 80% precision and recall in geospatial mapping of distribution grids.
• Demonstrated high generalizability, maintaining performance in Sub-Saharan Africa without retraining.
• Obtained an R² of 0.63 in estimating the fraction of underground power lines for wildfire risk assessment.

Conclusions:

• The developed framework offers a scalable and generalizable solution for mapping distribution grids using accessible data.
• The tool supports critical power system applications, including renewable energy integration and wildfire risk mitigation.
• This open-source framework facilitates the construction and maintenance of reliable and clean energy systems globally.