This is not merely a problem in theory. We have already witnessed several grid emergencies in regions with high renewables penetration (California Summer 2020, Texas Winter 2021) that occurred because a combination of weather events created generation shortfalls, resulting in brownouts or blackouts. While the underperformance of fossil-based resources was a major factor in these events, a significant driver of each episode related to the concurrence of extreme heating or cooling demand with declining levels of renewable generation.
Accurately characterizing climate and weather variability is key to understanding whether or not a utility or regional grid will be able to effectively meet customer demand over a multiyear or multi-decade planning horizon, a condition known as resource adequacy. Yet today, utility resource adequacy studies are generally built on a small sample of historical weather years–often generated from non-coincident temperature, wind, and irradiance data. This practice severely limits the accuracy and usefulness of such analyses.
In this work, we demonstrate the application of a stochastic climate model to the resource adequacy problem. The stochastic climate model generates a statistically valid, broad (1000 paths), coincident hourly weather data sample at population centers, utility-scale wind energy sites, and utility-scale solar sites in ERCOT over a 15-year period.
Each probabilistic weather outcome is correlated across time, location, and among all weather variables. Furthermore, the model adjusts hourly CDFs in future years according to observed long-term trends and the uncertainty in those long-term trends, embedding climate change signals as well as the uncertainty in those signals.
Building on these weather simulations, we calculate energy production estimates on each probabilistic weather path via machine learning models of hourly demand (as a function of temperature and dewpoint), hourly wind generation (as a function of wind speed and direction), and solar generation (as a function of irradiance and temperature). Each model is trained on representative, actual data from the ERCOT market.
Finally, we introduce a new metric we call “Grid CVAR” to estimate generation shortfalls. This measure conveys both the likelihood and magnitude of expected shortfall events, improving on the current Loss of Load Expectation metric, which only conveys the likelihood of a shortfall event occurring. We calculate Grid CVAR at several percentiles on each hour of the simulations and over rolling windows (4 and 8-hour). This facilitates actionable insights such as “there is a 1 in 20 chance of a 300 MW shortfall for hourly period X”, which grid planners may use as a roadmap to invest in mitigating resources such as battery storage or other controllable generation.

