7.4.1: Southern California Edison

Risk Analysis

In a discussion about risks, SCE stated that it used machine learning to determine the topographical and climatological risk factors used for short-term and current risk modeling. [1, page 17]

Equipment Analysis

In discussing the likelihood of equipment contributing to a wildfire either through operating or failure, SCE makes use af various sub-component models that use machine learning (ML) algorithms to assess the relevance of individual ignition drivers associated with the sub-component in calculating a probability of ignition (POI).

Sub-component models include:[1, pages 71-72]

• EFF: Equipment Facility Failure. These equipment models use historical equipment outage data (outcomes), current equipment condition attributes (e.g., age, voltage, inspection results, etc.), and relevant environmental attributes (e.g., historical wind, asset loading, number of customers, temperature, relative humidity etc.). The historical attributes and outcomes can be used as a training data set in machine learning.
• CFO-Veg: Contact from Object – Vegetation. SCE uses machine learning algorithms to develop predictive models for vegetation contact.[1, Table 5-1 SCE Risk Modeling Assumptions and Limitations, page 104]
• CFO-Other: Contact from Object – Other.

SCE uses machine learning algorithms to develop predictive models for equipment failure. [1, Table 5-1 SCE Risk Modeling Assumptions and Limitations, page 104]

Weather Forecasting

In a discussion about the use of weather data, SCE says that it uses machine learning algorithms to associate applicable weather variables from the weather research and forecasting (WRF) model at the time of fault/ignition events, using 2009-2023 weather data.[1, Table 5-1 SCE Risk Modeling Assumptions and Limitations, page 103]

The output from the weather prediction systems is used as as input to the machine learning modules. The machine learning modules and the forecast output is presented to Subject Matter Experts (SMEs). As of the end of 2024, SCE deployed machine learning to 1,624 weather station locations throughout SCE’s service area.[1, page 415]

SCE’s fuel moisture modeling makes use of SCE’s weather forecast output in conjunction with mathematical algorithms in order to estimate dead fuel moisture. SCE developed a machine learning model trained on historic weather and fuels data to predict live fuel moisture levels.[1, page 415]

SCE has five machine-learning-based weather forecast systems to augment traditional weather forecast output. SCE has developed machine learning models for up to 1,624 weather station locations using internal weather model data and up to 1,947 weather station locations from public forecast data.[1, page 416]

The initial machine learning forecast corrections have been expanded beyond the sustained wind speed and gust speed predictions to include temperature and dew point depression variables. Existing machine learning models have been retrained each year to learn from recent weather events. [1, pages 416-417]

University of California Santa Barbara (UCSB) also developed a deep learning model to translate irregularly spaced weather station observations into a regularly spaced grid with 500-m grid length, providing insight into the real-time weather conditions over locations that do not have weather station coverage. SCE is using UCSB's Deep Learning Gridding Meteorological Model (DLGNOME).[1, page 417]

SMART Meters

Like PG&E, SCE uses Smart Meters and can utilze data from them to detect evidence of transmission wires being down and can use data to help predict the need to replace transformers. For example:

• SCE's Meter Alarm Down Energized Conductor (MADEC) is a machine-learning (ML) algorithm that uses smart meter data to detect a subset of energized wire‐downs and other high-impedance faults/hazards. MADEC can generate an alarm that allows SCE to respond quickly to manually de‐energize the line or circuit.[1, page 396]
• SCE's Transformer Early Damage Detection (Transformer EDD) uses SMART meter data that is processed by a custom algorithm to proactively identify one failure mode of distribution transformers. Identified transformers can be scheduled to be replaced before potential failure in order to reduce safety hazards for the public, prevent grid disruptions, and power outages.[1, page 396] Winding shorts, partially turn-to-turn shorts, can create small increases in voltage on a transformer secondary that can be detected by SMART meters. By comparing voltage data of other transformers in the vicinity, SCE can create schedules for equipment replacement maintenance actions for some transformers prior to failure, thereby reducing ignition risks due to equipment failure and limiting the effects of electric service outages to customers.[1, page 678]

Smart meters collect a large amount of data. When coupled with other data, SCE can generate alerts for various inspections that might be needed.

Fire Detection Systems

SCE uses several types of fire detection systems:

• High Definition (HD) cameras and
• Satellite and other imaging technology.

References

1. Southern California Edison (SCE), 2026-2028 Wildfire Mitigation Plan, May 16, 2025, Revision 0. 2026-2028 Wildfire Mitigation Plan.