Loading...
Loading...
California's electricity system is arguably the most complex in the United States, and not primarily because of its size. The combination of CAISO's real-time energy market, the CPUC's resource adequacy (RA) program, the Low Carbon Fuel Standard's interaction with electrification load growth, and Pacific Gas and Electric's Public Safety Power Shutoff protocols creates a regulatory and operational environment where generic utility AI tools fail in specific, expensive ways. PG&E's PSPS program — which has de-energized transmission and distribution lines serving millions of customers during high fire weather events since the 2018 Camp Fire and Kincade Fire period — is itself an AI-informed decision process: PG&E now uses ML weather models, grid topology analysis, and customer impact scoring to decide which circuits to de-energize and in what sequence. The CPUC has scrutinized PG&E's PSPS methodology in ratable earnings proceedings and ordered third-party audits of the models used. Southern California Edison serves the Los Angeles Basin and Inland Empire under CPUC jurisdiction, with a different fire exposure profile concentrated in the Santa Ana wind corridor around the Palmdale and Rancho Cucamonga transmission infrastructure. San Diego Gas and Electric, the smallest of the three major IOUs, has the most advanced grid hardening program and was the first California utility to deploy AI-based situational awareness tools for fire weather management. Diablo Canyon Power Plant — the last operating nuclear facility in California, with two Westinghouse pressurized water reactors producing about 2,200 MW — has been approved by the California Legislature and NRC for a 5-to-10-year license extension beyond its original 2025 retirement date, creating a new market for nuclear AI maintenance tools at a facility that had been winding down its technology investment.
Updated June 2026
PG&E's Public Safety Power Shutoff program is operationally managed by a decision support system that combines National Weather Service forecast data, proprietary wind and humidity sensor readings from PG&E's own 1,300-plus weather station network, LiDAR-derived vegetation clearance data for 100,000-plus miles of transmission and distribution lines, and historical ignition data by circuit segment. The ML model that translates these inputs into circuit de-energization recommendations is subject to CPUC regulatory scrutiny — the Commission's Safety and Enforcement Division reviews PSPS events under rules established after the 2019 PSPS proceedings. AI vendors who want to work with PG&E on PSPS decision support need to understand that the output of their model will be reviewed by CPUC engineers and potentially scrutinized in PUC enforcement proceedings — model explainability and audit trail documentation are not optional features. SCE and SDG&E operate their own fire risk management systems; SDG&E's Wildfire Safety Operations Center in San Diego, which has been operational since 2017, is the most mature AI-augmented wildfire situational awareness platform among California's IOUs and has been studied by utilities in other fire-prone states as a reference architecture.
CAISO manages real-time dispatch for California's grid and operates the Western Energy Imbalance Market, which now covers most of the western US grid. Load forecasting in the CAISO footprint has to account for factors that no other US grid faces in the same combination: rooftop solar penetration that masks actual consumption demand, aggressive CPUC-mandated battery storage procurement that changes the net load ramp shape daily, the demand response behavior of industrial accounts under CPUC's Demand Response Program, and the interplay of the Low Carbon Fuel Standard with EV charging load growth that has added 2,000-plus MW of new evening demand in the Los Angeles and Bay Area metros since 2020. PG&E, SCE, and SDG&E submit load forecasts to CAISO under CPUC-approved methodologies, and forecast errors contribute to real-time ancillary service costs that ultimately hit ratepayers through the annual CAISO market results. CAISO itself runs ML-based load forecasting for ISO-level dispatch and has published research on its forecasting methodology. The CPUC's resource adequacy program — which requires LSEs to procure a 15% reserve margin of capacity above coincident peak load each year — uses utility load forecasts as the baseline, so forecast accuracy has direct capital allocation consequences: over-forecast and utilities build excess capacity; under-forecast and they face RA deficiency penalties.
Diablo Canyon Power Plant's approved license extension — enabled by California SB 846 in 2022 and subsequent NRC license renewal filings — has reversed a period of investment wind-down at the plant and created a new window for nuclear-grade AI predictive maintenance at a facility that had not anticipated continuing operations past 2025. PG&E's nuclear operations team at Diablo Canyon, operating under NRC license amendments, is evaluating AI condition monitoring tools for primary system components whose aging management programs were designed for a scheduled retirement that is no longer imminent. The specific applications are reactor vessel surveillance data analysis, steam generator fouling trend detection, and reactor coolant pump vibration monitoring — all areas where ML anomaly detection on existing instrumentation data can identify degradation trends without new sensor hardware. Nuclear-grade AI work at Diablo Canyon requires NRC-qualified software development processes (IEEE 7-4.3.2 for software quality), experience with PG&E's SONGS-informed plant programs, and the institutional patience to work within nuclear corrective action program timelines. Vendors with fleet-level nuclear experience — particularly with Westinghouse 4-loop PWR plants similar to Diablo Canyon Units 1 and 2 — are the realistic candidates. Ask any California nuclear operator and they'll tell you that the gap between a technically sound AI tool and a plant-approved tool is the quality assurance documentation stack, not the algorithm.
Connecting AI systems to existing business infrastructure and workflows
Workflow automation using AI, including Make.com-style automation and RPA
Predictive models, data analysis, and ML pipeline development
Image recognition, object detection, video analysis, and visual inspection systems
CAISO's RA program requires load-serving entities to procure capacity equal to their peak load plus a 15% reserve margin, verified annually through the CPUC's RA proceeding. Utilities that use AI forecasting to more accurately predict their coincident peak load can procure closer to actual need rather than adding conservative buffers that cost millions in capacity payments. PG&E, SCE, and SDG&E all maintain large forecast teams, and the difference between their AI-informed forecasts and traditional methods is now a material input to CPUC rate proceedings. The CPUC's Load Forecasting Proceeding (R.22-08-001) has examined forecast methodology in detail — utilities that can demonstrate AI-improved forecast accuracy receive more favorable treatment in RA compliance determinations.
PG&E's PSPS decision support system integrates weather forecasting, circuit risk scoring, and customer impact models. The weather layer uses ML models trained on PG&E's proprietary sensor network data alongside NWS forecast products. Third-party vendors have contributed to specific layers — fire weather forecasting companies like TomorrowIO and DTN have IOU contracts for enhanced short-range wind forecasting. The circuit risk scoring layer, which ranks which transmission segments are highest risk given forecast conditions, is largely internal PG&E intellectual property. New vendors entering this space typically start with a specific sub-problem — vegetation encroachment detection, conductor temperature modeling, or customer vulnerability scoring — rather than proposing to replace PG&E's integrated system.
The Low Carbon Fuel Standard creates a financial incentive for utilities to sell electricity to EV drivers — LCFS credits are worth $100–$200 per metric ton of CO2 avoided — which has accelerated utility EV charging program deployment. SCE's Charge Ready program has deployed 40,000-plus EV charging ports at commercial locations, and PG&E's EV Fleet program targets commercial fleets. The demand signature of this charging load is time-of-use sensitive and highly controllable, making it an AI demand response asset. But AI load forecasting that doesn't model EV adoption rates by ZIP code and charging behavior by customer segment will underforecast evening demand in the LA and Bay Area metro areas where EV penetration is highest. SCE's load forecasting team has published methodology papers on EV adoption modeling that represent the current California state of practice.
SDG&E's computer vision inspection program, which uses drone and helicopter imagery with AI analysis to flag vegetation encroachment and equipment anomalies on high-fire-risk circuits, has been the most-cited California utility AI program outside of PSPS. The CPUC has approved SDG&E's Wildfire Mitigation Plan, which includes AI-based inspection as a documented layer of fire risk reduction. PG&E has since deployed similar programs across its high-fire-threat district transmission lines in the Sierra Nevada foothills and North Bay. The cost benchmark for AI-augmented inspection programs in California runs $0.50–$1.50 per line-mile per year, depending on terrain difficulty and inspection frequency — compared to $5–$15 per mile for traditional helicopter survey methods.
Yes — and this is where California utility AI is most advanced globally. CAISO's grid now has over 10 GW of utility-scale battery storage under CPUC-ordered procurement, and optimizing that storage portfolio requires AI dispatch algorithms that balance energy arbitrage, ancillary services, and RA capacity value simultaneously. PG&E, SCE, and SDG&E each manage storage portfolios that require real-time dispatch optimization layered on CAISO's market signals. The commercially deployed platforms — including AutoGrid Flex, Fluence IQ, and Oracle Utilities' NMS extensions — have been validated in California conditions and represent the current standard of practice. New AI approaches being evaluated include multi-agent reinforcement learning for portfolio-level storage dispatch, which has shown 8–15% improvement over rule-based dispatch optimization in CAISO operating conditions.
Reach California businesses searching for energy & utilities AI expertise.
Get Listed