Machine Learning Helps Southern California’s Metropolitan Water District Maximize Solar Energy Investments


The Metropolitan Water District of Southern California (MWD) – which provides water to 19 million people in a 5,200 square mile service area – is using machine learning technology to improve the efficiency of its solar panels and quickly identify faulty solar panels. MWD, which has invested heavily in solar energy to power its facilities, has developed a deep learning model that automatically examines and analyzes thermal imagery of solar panels collected by drones flying over the department’s solar farms. This analysis can show areas of lower efficiency in solar farms and spot failing solar panels while still under warranty.


According to the district, automating thermal imaging analysis has significantly reduced turnaround time and improved the accuracy of this analysis. MWD staff first manually reviewed the thermal images, scanning for anomalies by hand and comparing those areas to high-resolution, natural-color images. The deep learning model can recognize the solar panels themselves and identify any thermal anomalies in the solar panels to the tolerances specified during model training. In an early proof-of-concept, the deep learning algorithm recognized verified solar anomalies at a rate nearly double that of a human analyst, MWD says.

MWD operations personnel use this data to quickly identify panel failures, obstructions, or other issues that could reduce efficiency, thereby protecting and maximizing the district’s investment in renewable energy. MWD now plans to expand its use of this technology to dam surface analysis, road pavement maintenance, crop fallowing, facility inspections and water usage. irrigated lawn.


Districts interested in taking advantage of this technology will need access to high-resolution natural color and thermal imagery, which can be captured by drone or aerial photography. If captured by an in-house drone program, the images should be processed by software capable of providing a spatially enabled mosaic that can be imported into a GIS software platform. A GIS analyst can use this mosaic dataset to identify features in the imagery that will be used to train the model. As with any machine learning algorithm, results will improve with additional practice and repetition.

Sherry J. Basler