This is the second blog post in our three-part series that explores key areas where Sunairio’s Omniscale Next-generation Ensemble (ONE) forecast model outperforms traditional solutions. Our first blog post demonstrated that Sunairio ONE was more calibrated, sharp, and extremes-conspicuous than legacy ensemble methods. Here, we show that Sunairio ONE’s high spatial resolution captures local variability of wind speeds better than existing models and demonstrate how that wind speed gradient can translate to a large variation in expected power output.

Introduction 

Variable renewable energy assets like utility-scale wind and solar farms experience meaningful fluctuations in power output as local weather conditions change. However, most weather forecasts aren’t generated at a spatial or temporal granularity that’s sufficient to accurately anticipate those dynamics.  In the case of wind, for example, detailed topographical features and their resulting phenomena (e.g., terrain-induced waking or speed-up effects) are lost at coarse resolutions¹. Sunairio ONE addresses these challenges by providing a high resolution weather forecast, which captures small-scale variability, enabling best-in-class asset-level generation potential forecasts.

Wind farm forecasting best practices   

Asset-level wind energy forecasting requires both highly-accurate weather forecasts of mesoscale wind (on the order of a few km) and sophisticated models of wind field dynamics within the wind farm footprint, which are affected by many factors including terrain, vegetation cover, and turbine waking.

In fact, we find that wind energy modeling best practices dictate separating this problem into three stages: 1) the mesoscale free-field wind² forecast, 2) the microscale wind field within the wind farm footprint (less than 1km effects usually influenced by terrain), and 3) turbine-by-turbine waking (Figure 1).

Figure 1. Diagram showing three mains steps involved in accurate modeling wind energy at a wind farm. 

Sunairio has already pioneered novel solutions to Steps 2 and 3, previously reporting on our use of highly-efficient CFD models to model turbine-scale (100m) wind fields and waking for all operational wind farms in the U.S.

We are now excited to leverage Sunairio ONE, our high-resolution ensemble weather forecast, to improve mesoscale wind forecasting for asset-level generation.

Existing weather forecasts are too coarse, or fine-but with limited outlook    

Surveying the landscape of publicly available weather models that can provide mesoscale free-field wind forecasts shows that none offer truly high spatial and temporal resolution output that extend beyond a few hours (Table 1). Furthermore, Figure 2 helps provide some scale to the current problem, comparing the spatial resolution of Sunairio ONE (at approximately 2km) and the IFS (at approximately 10km) against actual wind farm footprints in ERCOT.  Mesoscale forecasting at 10km (like the IFS) vs. 2km may average out important wind gradients that vary over a wind farm and lead to exponential wind generation errors (more on this below). 

Additionally, Sunairio ONE forecasts are generated at hourly temporal resolution–critical for energy forecasting applications–for the full forecast period; no other major publicly available forecast model does this. For example, the IFS starts at hourly resolution for the first three to four days but then drops to 3-hourly and 6-hourly steps as it gets further out.

Forecast Model	Outlook Period	Temporal Resolution	Spatial Resolution
HRRR (NOAA)	48 hours3	1-hourly	~	3km
GFS/GEFS (NOAA)	16 days	0-120 at 1-hourly 123-384 at 3-hourly	0.25°	~28km
IFS/IFS ENS (ECMWF)	15 days	0-90 at 1-hourly 93-144 at 3-hourly 150-360 at 6-hourly	0.1°	~10km
AIFS/AIFS ENS (ECMWF)	15 days	0-360 at 6-hourly	0.25°	~28km
Sunairio ONE	15 years	1-hourly	0.02°	~2km

Table 1. Comparison of forecast models by outlook period, temporal resolution, and spatial resolution. 

Figure 2. Zoomed area of map of wind farms in Texas (shaded bounding boxes show footprints of wind turbines at wind farms). 

Wind is highly variable across wind farm scales

To visualize just how important it is to capture variations in free-field wind forecasts, we analyzed a day of wind forecast performance (three days out) between Sunairio ONE and the IFS across a major ERCOT wind farm, Capricorn Ridge. As Figure 3 shows, the difference between wind speeds at the north end of the farm compared to the south end (turbines marked in red in left panel) were as great as 3 m/s in some hours. While Sunairio ONE successfully forecasted this wind field gradient (right panel), the lower-resolution IFS could not resolve the proper dynamics, instead seeing a fairly uniform wind field.  

Figure 3. (Left panel): Map of Sunairio ONE forecasted wind speed across a wind farm in Texas over a 24-hour period; (Right panel): Line plot tracking the wind speed gradient (i.e., difference in wind speeds) from the north side to the south side of the farm compared toSunairio ONE 3-day ahead forecast and IFS ENS 3-day ahead forecast.

Why it all matters: Even seemingly minor weather variations can have outsized impacts in power output

Does a 3 m/s gradient in free-field wind speed over a wind farm really matter? Depending on the absolute wind speed levels, it can matter immensely. Below their rated capacity, wind turbine power output scales cubically with wind speed which causes even seemingly minor variations in wind speed to lead to large differences in generation potential. Turning to Figure 4, we plot an example power curve of a 1.5 MW turbine, which represents the majority of turbines at Capricorn Ridge. A difference of 3 m/s in wind speed can translate to up to a 959 kW delta in power output, or 64% of rated capacity!

Figure 4. Power curve for the 1.5 MW wind turbines that comprise the Capricorn Ridge Farm. 3 m/s wind speed gradients can translate to power differences as large as 959 kW.

Over a wind farm with hundreds of turbines, the impact of modeling wind speeds at a lower resolution on generation potential forecast performance can quickly add up.

Conclusion

Accurately forecasting local wind speeds is essential for creating reliable forecasts of asset-level generation. Sunairio ONE provides 2km spatial resolution, hourly temporal resolution forecasts that are ideally suited to serve as the critical free-field wind forecast step in wind farm generation modeling pipelines.

1. Energy Systems Integration Group. 2023. Weather Dataset Needs for Planning and Analyzing Modern Power Systems (Full Report). A Report of the Weather Datasets Project Team. Reston, VA. https://www.esig.energy/weather-data-for-power-system-planning ↩︎
2. Free-field wind speeds are wind speeds expected in open, unobstructed areas free of turbulence caused by structures such as wind turbines. ↩︎
3. The HRRR goes out to 48 hours only for the 0Z, 6Z, 12Z, and 18Z initializations. All other initializations go out 18 hours. ↩︎

This is the first edition in a series of three blog posts that explores the commercial shortcomings of legacy forecast solutions compared to the advantages of Sunairio ONE.

Introduction: Why the Power Sector Needs Ensemble Forecasts

While every business relies on forecasts to some extent, power sector businesses are forecast super-consumers, especially of weather and weather-based forecasts for power demand, wind generation, and solar generation. In fact the growth of renewables has strengthened the link between weather and energy, introducing weather variability to the supply side of the power grid supply-demand balance. This exponentially increases both the requisite forecast complexity of everyday grid operations and the manifest importance of reliable forecasts themselves.  

Regardless of their specific role, power-sector forecast consumers fundamentally make quantitative business decisions that rely on weather forecasts. Yet this business need runs straight into two of the most confounding aspects of weather and weather forecasts: the inherent volatility of weather itself and the large uncertainty of weather forecasts. 

Most sophisticated weather forecast users have long recognized that the best-practice solution to this problem is to not rely on one forecast, but many. An ensemble approach to forecasting involves issuing many alternative forecasts for the same phenomenon. A common use case is the ubiquitous hurricane track forecast plot that creates an ensemble of many possible trajectories to convey a sense of probabilistic forecast tracks. There are many ways to generate a forecast ensemble and the concept isn’t limited to weather forecasting, but it is widely accepted to be one of the best methods for investigating forecast uncertainty.

Creating an ensemble weather forecast theoretically satisfies two important needs for decision-makers: 1) it theoretically describes the range of possible outcomes, and 2) it can be used to calculate the implied probabilities of all outcomes within the range. 

Properly seeing the full range of outcomes – including extreme events – is absolutely critical in power grid applications because extreme events have an outsized effect on reliability and economic outcomes. For example, in the ERCOT real-time power market, just 1% of hours per year account for one-third of total annual market value¹. This means that having a clear view of tail risks is paramount for any power grid planner or power market participant, lest they make decisions with a limited forecast set that hides massive risk.

Moreover, it’s important to note that in the context of the modern power grid, extreme events aren’t limited to things like severe storms. Unusual combinations of weather that increase demand while reducing renewable energy generation (such as summer heat that lingers through twilight) is every bit as consequential from a reliability or market perspective.

In this blog post, we evaluate how successful traditional ensemble forecasts and the new Sunairio ONE ensemble are at generating reliable and actionable predictions for power grid/power market applications. We evaluate these ensemble forecasts in three dimensions: calibration (having correct probabilities), sharpness (seeing events clearly), and a new ensemble forecast performance dimension we introduce here: the principle of conspicuous extremes (seeing extreme or tail events).

Classes of Ensemble Weather Forecasts 

Ensemble weather forecasts come in many flavors, from the simple historical analogue (a set of historical observations, like historical weather years)  to compute-intensive traditional numerical weather prediction (NWP) algorithms, to highly complex deep learning-based generative AI models.

Historical Weather Analogue Ensembles

An historical analogue ensemble of, say, 30 years, constructs a thirty member ensemble where each member of the ensemble is a year of observed weather in the last 30 years. As Historical weather analogues rely on weather observations, serially-complete hourly analogues are generally limited to 75 years or less. 

Numerical Weather Prediction Ensembles

Starting from real atmospheric observations, NWPs simulate physical dynamics of the atmosphere to generate a model of future weather. Due to the complex and chaotic nature of atmospheric fluid dynamics, small perturbations in initial weather conditions can lead to dramatically different weather paths. NWP ensembles can take advantage of this by perturbing the initial conditions to simulate an ensemble of “representative” weather realizations. The IFS Ensemble (IFS ENS), an NWP forecast produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), is the gold standard against which all modern weather forecasting systems are compared due to its relatively sharp, skillful short-term forecasts. The IFS ENS consists of one control forecast and 50 additional ensemble paths.

Generative AI Weather Forecast Ensembles

Advances in deep learning have spurred a new paradigm of weather forecasting that borrows ideas from generative AI to produce global weather forecasts that are (1) competitive against the IFS Ensemble and (2) far more computationally efficient. Even without knowing any physical laws, these models, trained on historical reanalysis weather data, produce physically realistic weather patterns. The ECMWF has released an operational ensemble AI weather forecast, the AIFS Ensemble.

Ensemble Performance Metrics: Calibration, Sharpness, and the Ability to See Extremes

The taxonomy of ensemble forecast methods is broad, but many suffer from the same problems: poor calibration, lack of sharpness and poor visibility of extreme events. All of these dimensions are critical for a power sector forecast consumer. A sharp but uncalibrated ensemble will be overly confident but often wrong. A skillful calibrated ensemble that doesn’t satisfy the principle of conspicuous extremes will leave a forecast user exposed to lurking tail risks; you can’t plan for what you can’t see.

Figure 1 illustrates these dynamics. Starting from the left, panel A shows a hypothetical ensemble forecast that is calibrated but not sharp (its implied probabilities are correct but its poor skill means that the ensemble range is very wide), panel B shows an example of an ensemble that is sharp but uncalibrated, while panel C presents an ensemble that is both sharp and calibrated but completely missed an event outright because it does not model extremes. The ensembles in the first two panes have very obvious deficiencies while the ensemble in the third pane exposes the forecast user to more insidious risks – the forecast looks reasonable, performs well in normal weather conditions, but quietly hides tail risks.

Panel D, on the other hand, depicts an ideal ensemble forecast: sharp, calibrated, and extremes-conspicuous.

Figure 1. In the first three panes we show three hypothetical 50 path forecast ensembles that each suffer in one of the three forecast performance dimensions: calibration, sharpness and the principle of conspicuous extremes. In the third, we show an ideal ensemble with 10x as many paths. It is just as sharp as panel C but with more paths, capturing the tails of the distribution.

As we show in this blog post, legacy ensemble forecasts commonly exhibit the deficiencies shown above in A-C, while the new Sunairio ONE method is intelligently designed to replicate the ideal ensemble paradigm in D. 

A New Paradigm for Ensemble Forecasting: Sunairio ONE

Created in response to the significant limitations of the traditional ensemble forecast classes, Sunairio ONE is powered by a hybrid forecast architecture and designed to maximize commercial actionability. The ONE forecast begins by creating a 1,000-path weather ensemble via a long-range generative weather algorithm that samples from known distributions. Within the medium range window (1-14 days), a deterministic Control forecast is then created by solving an optimal forecast combination problem given the most recent public model guidance. Finally, the original seasonal ensemble is reshaped – conditional on the Control forecast – such that the expected ensemble forecast error matches the ensemble spread.

In the following sections we compare the ensemble forecast of Sunairio ONE to benchmark models in applications for both short and long range forecasting. 

Empirical Ensemble Model Performance Analysis 

Long-range Forecasts

Over longer time horizons (months to years), use of the Historical Analogue approach for ensemble forecasting in the power sector is commonplace as there are no skillful hourly weather forecasts. Accordingly, we compare how well Sunairio ONE performs over a long lead time by examining its performance against a 74-year Historical Analogue. As Figure 2 shows, we plot the ensemble calibration of Sunairio ONE against the Historical Analogue for a 12-month forecast that was issued with 6-18 months forecast lead time, using Daily Maximum Temperatures at Philadelphia airport (PHL). This type of plot, known as a probability-probability (PP) plot, compares the implied ensemble event probability (x-axis) to the actual observed event frequency (y-axis). As the plot shows, a 74-year historical ensemble is poorly calibrated (note how it generally underpredicts the probability of daily maximum temperatures– is biased low–a result of not adjusting for climate change), while the Sunairio ONE ensemble closely hews to the 1:1 line that represents good calibration: implied ensemble forecast probabilities = observed frequencies.

Figure 2. PP plot of 74-year Historical Analogue ensemble and Sunairio ONE for daily maximum temperature for the period July 2024 to June 2025. Historical Analogue for the years 1950-2023. Sunairio ONE trained through 2023 and then predicted for the Jul-24 to Jun-25 period (6 to 18 months forecast lead time) 

The Historical Analogue approach also performs poorly when considering its ability to see extremes. As Table 1 shows, the 74-year Historical Analogue completely missed 184 hours per year within the analogue range (a forecast surprise), as there are simply too few samples (74) to accurately represent the range of likely weather over an 8,760 hourly period. By contrast, Sunairio ONE captures 97% of those missed extreme values (178 of the 184 misses) due to the more robust 1,000-member ensemble that’s capable of seeing well past 99th percentile events.

	# of Hourly 2m Temperature Surprises, PHL July 2024 - June 2025
74 Year Historical Analogue (1950-2023)	184
Sunairo ONE (6-18 months out)	6

Table 1. Hourly 2m temperature surprises at PHL for the period July 2024 to June 2025 from two different approaches to long-range ensemble forecasting

Short- and Medium-range Forecasts

Over shorter time horizons we compare Sunairio ONE to benchmark medium-range (14-day) public ensemble forecast models. As Figure 3 shows, Sunairio ONE again outperforms traditional benchmarks such as the ECMWF’s IFS Ensemble by generating a calibrated forecast – even through the upper quartile event range (events in the 75th to 100th percentile). As we see, Sunairio ONE replicates extreme events at the correct observed frequencies while the IFS Ensemble (the gold standard NWP ensemble) drastically underpredicts them.

For example, the dotted green line indicates that temperatures predicted by the IFS Ensemble to occur no more than 5% of the time (the 0.95 non-exceedance probability level) actually occurred more than 21% of the time (a non-exceedance frequency of about 0.79), meaning that the these temperature events were about 4 times more likely than the IFS ensemble implied.

Figure 3. PP plot of Sunairio ONE and the IFS Ensemble for hourly temperature at PHL. PP plot is zoomed in to the upper quartile. Dotted green line shows that the IFS Ensemble p95 level is actually the observed p79 frequency.

Moreover, Sunairio ONE isn’t just calibrated – it’s also demonstrably sharper than traditional ensemble forecasts. Comparing the continuous ranked probability score (CRPS – a metric used to evaluate the accuracy of a probabilistic forecast) of Sunairio ONE to the GFS Ensemble (known as GEFS, NOAA’s benchmark medium-range forecast ensemble) we see that Sunairio ONE improves CRPS across a suite of weather variables by 20% or more (Figure 4).

Figure 4. Relative improvement in CRPS for Sunairio ONE compared to GEFS across several weather variables and daily aggregations.

The Sunairio ONE ensemble also does a much better job at seeing extreme events than traditional forecasts. Here we note that traditional weather forecasts actually tend to be worse at seeing extremes the closer they are to happening. For example, Figure 5 plots the number of hourly temperature forecast surprises (events in which the actual temperature realized completely outside of the ensemble range) for the IFS Ensemble at PHL over a period from June 20 to September 22, 2025 – with just one day of lead time. That is, we collected IFS Ensemble temperature forecasts for the following calendar day every day during this period and counted the number of hours in which hourly temperatures realized above or below the 51-member ensemble set.

Over this 95-day period, there were a total of 412 hourly forecast surprises (including misses above and below the ensemble range). Notably, the IFS Ensemble was nearly perfectly unbiased during this period at PHL, with MBE measuring +0.02 deg F, meaning the misses were not caused by the ensemble range being systematically shifted up or down. Rather, the forecast was simply overconfident (and wrong). We do not show the Sunairio ONE forecast surprises in this plot because the Sunairio ONE ensemble did not miss any hours.

Figure 5. Daily counts of hourly forecast surprises (hourly realizations outside the 51-member ensemble range) for the IFS Ensemble, June 20 to September 22, 2025 at PHL, with one day of forecast lead time. 

Next we measure the forecast surprises on daily maximums and daily minimums for 1 day, 7 days, and 14 days forecast lead time and compare them to the number of forecast surprises that should be expected given the ensemble size. Table 2 presents the results. As the table shows, the IFS Ensemble missed 3 daily max/min events at 14 days and 8 daily max/min events at 7 days (compared to an expected range of 2 to 6), but missed a whopping 17 days (18.5% of the analyzed period) at just 1 day of forecast lead time. This data suggests that the IFS Ensemble has an inherent underdispersion weakness which becomes more profound as the forecast lead time decreases. 

To estimate the magnitude of the underdispersion, we calculate how much wider the ensemble would have to be so that the actual misses fall within the expected range. As the table shows, the underdispersion is at least 1 deg F at 7 days of lead time and more than 3 degrees at 1 day of lead time.

Lead	Sunairio ONE exp. # surprises	Sunairio ONE # surprises	IFS ENS exp. # surprises	IFS ENS # Surprises	IFS ENS Underdispersion
1 day	0 to 1	0 (0%)	2 to 6	17 (18.5%)	> 3F
7 days	0 to 1	0 (0%)	2 to 6	8 (8.6%)	> 1F
14 days	0 to 1	0 (0%)	2 to 6	3 (3.3%)	0F

Table 2. Expected forecast surprises, actual forecast surprises, and estimated underdispersion for Sunairio ONE and the IFS Ensemble. 2m temperature daily maximum and daily minimum forecasts at PHL, one day forecast lead time, June 16 to September 28, 2025.

Real-world Implications of Bad Ensemble Forecasts

Forecast surprises such as these have immense consequences for power grid operators and power market participants because they represent hidden risks that generally haven’t been accounted for in any decision-making. That is, traditional ensemble weather forecasts, especially those issued for the next calendar day, are commonly thought to encapsulate a complete range of risk–but they do not.  

In fact as the left column of Figure 6 shows, during a three-day period from September 26 to 28 this year the IFS Ensemble range missed each day’s maximum temperature just one day out. There are 51 ensemble members within the IFS ENS and not a single one was high enough. 

Comparing that performance to Sunairio ONE (right column of Figure 6), we see that while the realized temperature was indeed towards the upper range of our one-day forecast, it was well within the ensemble spread–visible to any forecast user making quantitative scheduling, dispatching, or trading decisions.

Figure 6. One day lead time 2m temperature forecast ensemble range for IFS Ensemble (left) and Sunairio ONE (right), along with observations. PHL airport, September 26-28, 2025.

As these plots show, Sunairio ONE ensembles exhibit the overall characteristics of the ideal ensemble forecast from Figure 1: calibrated, sharp, and showing the risk of extremes.

To put the IFS Ensemble forecast surprises in context of power grid balances, we convert the temperature misses seen here into hidden PJM load risks (assuming a similar temperature surprise across the PJM footprint). Figure 7 plots the result, showing that a ~2-4 deg F temperature miss across these hours accounts for up to 6 GW of hidden load risk–equivalent to the regional power demand of the entire Philadelphia metro area!

Figure 7. Magnitude of potential hidden PJM load due to IFS Ensemble temperature surprises, September 26-28, 2025, and daily max load for PECO (Philadelphia Electric Company).

This means that a major city’s worth of demand risk was hiding in plain sight, invisible to forecast users because the benchmark models fail to see extremes – even just one day out. 

Conclusion

Ensemble forecasts are critical tools for grid operators, power market participants, and other decision-makers in the energy sector. However, legacy ensemble forecast methods suffer from chronic and significant deficiencies that obscure important risks: missing hundreds of hours per year, misrepresenting event probabilities by a factor of 4, and hiding extreme events equivalent to a city’s-worth of power demand. 

Sunairio ONE addresses these problems by intelligently generating an ensemble forecast that’s calibrated, sharp, and sees the all-important extreme events.

1. Data for ERCOT North Hub RT, 2019-2023 ↩︎

Sunairio ONE captures 97% of extreme events that traditional forecasts miss, providing critical intelligence for grid reliability, power systems planning, and energy markets

Baltimore, MD. — October 16, 2025 — Sunairio, pioneer of next-generation grid forecasting software that provides critical weather and energy insights, announced today that it has launched Sunairio ONE (an Omniscale Next-generation Ensemble). Sunairio ONE now powers the company’s award-winning software for modern energy risk management.

Anticipating granular grid risk today is a daunting task that’s frustrated by a disjointed and inadequate technology landscape. Conventional weather forecasts are complex and slow, limiting their usefulness to modeling near-term events over relatively large areas — with forecast scenarios that misrepresent the chance of extremes. Longer-term forecasts are so inaccurate that the meteorological agencies themselves warn against using the data directly, forcing the industry to rely on historical analogues instead. And the new AI weather forecasts are unfortunately linked to these problems because they train off the same old data. 

Sunairio ONE makes these approaches obsolete.

“We built Sunairio ONE from the ground up to solve these challenges, enabling unparalleled resolution, accuracy, and correct probabilities, complete with seamless coverage for the next hour, the next 15 years — and anywhere in between,” said Rob Cirincione, CEO, Sunairio.. “The key is our unique data — Sunairio ONE is trained on the Sunairio High-resolution Earth Dataset (the SHED), our proprietary, high-resolution historical weather data archive that’s purpose-built to replicate granular risks for the modern grid.”  

Sunairio started with the industry’s benchmark historical weather data, leveraged machine learning to sharpen resolution by 100x, then trained ONE on that data while incorporating changing climate fundamentals. Sunairio ONE generates 1,000-member hourly ensembles compared to traditional methods that stop at 50. The result is unprecedented forecast fidelity for load, renewables, and grid stress across all time scales.

“Alternative forecast methods rarely issue more than 50 forecast scenarios, preventing users from seeing the make-or-break events,” Cirincione explained. “Imagine if I told you there was a 1-in-50 chance of winning the lottery, except you don’t win the lottery, you black out the power grid or bankrupt your company. That’s the risk we’re taking with these limited forecasts.”

“More recently, there’s been a burst of new AI-powered weather forecasts, but what often gets overlooked is that these AI models are only as good as the data they’re trained on — and these models have all been trained on the same public, relatively low-resolution archive,” continued Cirincione. “That means they’re all susceptible to missing the same events — especially extreme events — which are critical in modern power grids where just 1% of hours account for 30% of grid stress.”

A larger forecast ensemble also improves accuracy. For example, compared to grid-level forecasts provided by ISO operator ERCOT, Sunairio’s hourly wind and solar forecasts are up to 20% more accurate on average while capturing the true risk of forecast misses and unprecedented extremes in the ensemble distribution.

Sunairio ONE is an evolution of the company’s award-winning, long-term generative forecast technology — previously recognized by the National Science Foundation, American Clean Power Association, and EPRI. The company’s high-resolution, large-ensemble approach to forecasting now spans both seasonal and climate scales (from 15 days to 15 years) as well as shorter-term 1- to 15-day forecasts.

"No other solution gives us weather and energy forecasts from a 1,000-member ensemble and continuously tracks changes to load growth and infrastructure buildout," said Josh Henson, Vice President, Wholesale Trading, Constellation. “We’ve been pleased with Sunairio as a platform to model probabilistic weather and energy risks for longer-term power markets. Sunairio ONE provides those same insights for short-term markets enabling better business decision-making that benefits our customers.”

Sunairio ONE forecasts support actionable intelligence for reliable planning across power markets and utility operations. It is designed to support a range of use cases, including:

• Probabilistic energy trading and hedging strategy: Ensembles predict future energy price ranges and show potential for risk versus reward.
• Asset-level renewables forecasting: Sunairio ONE provides probabilistic forecasting for individual renewable energy assets (including availability and curtailment losses) for independent power providers, utilities, and public market participants.
• Climate-aware portfolio risk management: Accurately calculates the risk to energy investments resulting from weather-driven impacts across load, wind, and solar.

“Some of the greatest risks to power grids are hiding in plain sight. They don’t always look like hurricanes, which makes them more insidious. Increasingly, they look more like unusual combinations of weather that increase demand while reducing renewables output: heat that lingers past twilight or polar vortex cold snaps that arrive with calm winds,” said Raiden Hasegewa, PhD, Director of Data Science, Sunairio. “By combining weather and energy in a sophisticated forecast model, we’re giving companies insight they can’t get anywhere else.”

All of Sunairio’s energy and market models built on top of Sunairio ONE automatically update and retrain continuously, keeping pace with fast-moving fundamentals. For more information, please visit sunairio.com or email info@sunairio.com.

###

About Sunairio
Founded in 2020, Sunairio is the pioneer of award-winning, next-generation grid forecasting software that’s the first to provide integrated energy, weather, and climate insights. Sunairio helps energy traders, grid operators, utility-scale asset developers, and VPP and demand response aggregators make better commercial decisions in the face of increasing grid variability and extreme event risks. Sunairio and Sunairio ONE have received recognition from the NSF, ACP, and EPRI. For more information, please visit sunairio.com.

Media Contact
Nikki Arnone, Inflection Point Agency for Sunairio
nikki@inflectionpointagency.com

Late last month, on June 24, 2025, PJM experienced a record-setting heatwave that sent heat indexes soaring above 100^oF throughout the Mid-Atlantic, hotter than anything observed in late June since at least 1950. This extreme weather event sent regional power demand and prices spiking — though not at the same time. 

The price spike occurred in the two hours after demand peaked that day — a (perhaps unexpected) consequence of the growing reliance on renewables in PJM. Here's how solar fundamentally changed PJM's risk profile on that sweltering June day.

How net load drives grid stress and power prices

Net load — not native load — drives today’s price spikes. PJM native load peaked in hour ending (HE) 18, but Western Hub LMP spiked to $1,500/MWh+ in subsequent HE 19 and 20, as the plot in Figure 1 shows. 

Why? Because those are the hours in which net load peaked. As we’ve discussed before, net load (native load minus renewables) is the primary driver of grid stress in power markets with a significant share of renewables. And when grid stress rises, so do prices.

We can now officially count PJM as a market in which renewables can’t be ignored.

As Figure 2 shows, there was approximately 10 GW of solar generation in PJM on June 24 during hours 10–17, which significantly reduced net load — and therefore grid stress — throughout the day. Even in the peak load hour (HE 18) there was still 9 GW of solar generation. 

But this solar output dropped precipitously as the sun set across the ISO footprint to just 3 GW in HE 20 — causing net load (and overall grid stress) to continue to rise even as native load was falling. The resulting net load ramp required PJM to quickly dispatch expensive units to stabilize the grid.

The end result? The highest prices we’ve seen in PJM since 2022 occurred much later in the day than would have been historically expected for this period in June.  

Solar is shifting PJM fundamentals

Peak net loads have been creeping later in the day for some time. The highest net load hour in June has shifted from HE 17 in 2018 and 2019 to HE 20 this year, as Figure 3 shows  for each June’s highest net load day since 2018.

This pattern extends to less extreme days as well: 70% of daily net load peaks occurred at hour ending 19 or later in 2025 compared to 0% of peaks occurring that late in June 2018, as Figure 4 demonstrates across all June days since 2018.

This intraday net load pattern shift mirrors the massive growth of installed solar capacity in PJM. Utility-scale solar capacity has increased roughly 500% since 2020, from 2 GW to more than 12 GW as of June 2025, as Figure 5 illustrates.

As solar penetration increases in PJM, it drives greater price volatility and concentrates the highest price risks later into the evening hours during summer months. 

Fundamentally, it creates a grid management challenge as the quickly setting sun necessitates a steep ramp for other generators to be brought online. This highlights the need for accurate net demand (load, solar, and wind generation) forecasting and fast-ramping resources to successfully enable renewables integration.

Even with the headwinds that new solar projects face today, we still expect a sizable capacity addition over the next few years as projects that are under construction reach completion and the pipeline of projects that qualify for remaining tax incentives break ground. 

The bottom line: add PJM to the list of power markets whose fundamental outcomes are increasingly controlled by the intraday volatility and intermittency of weather-driven generation resources.

In modern electric grids, grid operators must continually balance power demand with power generation to maintain steady grid frequency and ensure overall grid stability. This entails anticipating, in real time, short-term weather-driven changes in both customer load and renewables output to match the remaining “net demand” with dispatchable generation. Grid planners, on the other hand, look far into the future to ensure long-term “resource adequacy” — i.e., that future generation resources will be sufficient to cover demand under all possible weather scenarios and grid conditions.

Satisfying resource adequacy through long-term grid planning processes is a critical component of grid reliability because generation resources and transmission infrastructure can’t be built in the few days ahead of a forecasted heat wave or winter storm. 

To study long-term resource adequacy, grid planners:

1. Construct a set of future weather scenarios
2. Make some assumptions about asset-level availability
3. Calculate the resulting hourly net demand given anticipated customer load projections and renewable energy capacity buildout
4. Quantify the likelihood of capacity shortfalls that would necessitate shedding load to maintain grid stability (i.e., brownouts or blackouts)  

Resource adequacy studies, in other words, rely on an accurate characterization of future weather variability. In practice, grid planners typically use historical weather as a proxy for future weather — tacitly assuming that a) future weather will be similar to historical weather and b) historical weather is a large enough sample to assess the risk of extreme events.

Unfortunately, both of these assumptions are wrong. Future weather is different from historical weather. Average temperatures, wind speeds, and irradiance are all expected to shift over time from climate change effects, to varying degrees and in different directions depending on what region is being studied. Moreover, the shape of weather distributions will also change. For example, climate modeling shows that extreme cold risks may actually become more frequent in the future as climate change weakens the jet stream — even though average temperatures will rise. That means that the bottom tail of a temperature distribution gets longer while the rest of the distribution shifts right.

Regarding the second assumption — that we have enough historical weather data to assess extremes — the historical record is much smaller than you might think.

First, we can only use serially-complete historical datasets that have values for every hour of the year because we can’t make reliable statistics from partial-year data. This generally excludes airport station temperature data before 1980.

Next, we have to find high-quality weather data to model wind and solar generation. Traditionally, that has meant relying on satellite-based irradiance data (available since 1998) and the NREL WIND Toolkit, which is a high-resolution wind speed dataset calculated for 2007-2013.

As a result, a resource adequacy study that includes correlated temperature-driven load, irradiance-driven solar generation, and wind-driven wind generation would only be able to call upon 7 historical year samples — drastically limiting the view of weather extremes that the grid could face. 

Sadly, flawed grid planning due to insufficient weather assumptions can lead to catastrophe, with major power outages occurring recently in California (2020), Texas (2021), and North Carolina (2022). An NREL post-mortem of all three events found modern electric grids, which increasingly rely on intermittent weather-dependent renewable energy generation, to be increasingly impacted by extreme weather conditions — and that “weather in recent years has exceeded the bounds of anticipated conditions,” highlighting the need for improved planning processes that accurately account for jointly correlated extremes of weather, power generation, and generation outages.

To fill this gap, Sunairio generates a 1,000-path ensemble of future hourly weather to give grid planners robust, complete, and actionable distributions of weather, grid conditions, and asset-level variability. This ensemble is climate-change adjusted, trained on the longest possible series of weather data, and numerous enough to gain intelligence into future extreme events — in whatever form they may come.

For plants, spring is a season of growth. But for power market participants and grid operators, it’s a time of surprises and volatility.

While typically mild spring weather results in low expected grid demand, extreme spring weather — through a confluence of factors — can drive real-world hourly grid balances to levels that approach or exceed emergency conditions. This contrast between mostly moderate days and acute periods of serious grid stress makes spring just as challenging to navigate as the traditional “peak” seasons in summer and winter.

Properly anticipating this inherently stochastic risk requires both a nuanced understanding of the underlying fundamentals and a probabilistic framework to quantify low-probability but high-impact economic and reliability outcomes. In this blog post, we examine actual weather, grid, and market events from this spring in ERCOT within a probabilistic context. Looking ahead to spring 2026, we then explore how rapidly changing grid fundamentals may alter next year’s spring grid risk profile.

Late season cold and an early heat scare in Texas

This year has not been an exception to the rule that spring weather varies wildly. Temperatures on April 7 dropped to 40ºF at DFW and into the 30s around Dallas (colder than all of March) while an early season heat wave drove forecast highs over 100ºF last week (levels not typically seen until July).

Figure 1 plots this spring’s weekly average, minimum, and maximum temperatures (green lines) against ranges derived from Sunairio probabilistic climatology (box and whisker plots) at DFW. As the plot shows, most hourly temperatures throughout spring are relatively mild — between 50ºF and 70ºF — but extremes dip into freezing territory and extend into severe heat.

An epic forecast fail

Temperatures could have been even more extreme this year if the weather forecasts for the May 12 week hadn’t flopped. As we see in Figure 2, the forecast for the May 14 afternoon high at DFW was 102ºF just 12 hours out — yet realized 8 degrees lower at 94ºF. That 8-degree temperature forecast error likely reduced ERCOT RTO load by approximately 9 GW versus higher temp load expectations, leading to a $40 Day-ahead/Real-time spread (DA higher than RT) — highlighting the difficulties of navigating grids amid temperature variability and forecast uncertainty.

Generation outages peak

In spring, dispatchable resources are often not available on purpose. Given that the majority of Texas spring weather is mild and that the season immediately precedes the peak demand summer months, generators schedule the bulk of their maintenance outages during this time. As we see in Figure 3, nonrenewable (thermal) generation outages usually peak in early April (coinciding with the mildest expected temperatures and lowest expected load), though unscheduled outages can cause significant variability. Ironically, such high levels of dispatchable unit outages can tip the grid from normal conditions into capacity shortfalls. Scheduling generation outages in spring makes sense, until it doesn’t.

Renewables ramp up

Temperatures and unit outages aren’t the only thing rising in ERCOT during spring. Wind generation typically maxes out in early April and solar generation typically increases by 38% from March 1 to May 31. Further complicating grid dynamics are intraday generation patterns, where wind generation peaks overnight and solar generation is dictated by the diurnal sunup-to-sundown cadence. Figure 4 plots both these seasonal (left panel) and intraday (right panel) trends.

Implications for power markets and reliability

Understanding grid risk these days is much more than understanding a simple temperature -> load relationship. It’s a multidimensional puzzle whose solution is driven by correlated variability between load, wind generation, solar generation, and the availability of dispatchable resources. It’s sensitive to temperature extremes, shifts in renewable generation patterns, and unit outages.

At Sunairio, we combine these fundamentals into one metric that provides a reliable indicator of overall grid stress — the Grid Stress Index (GSI), which measures the ratio of Net Demand (load minus renewables) to Available Dispatchable Capacity (read more about the construction of GSI here).

As Figure 5 shows, this metric anticipates hourly price volatility, with conditions in ERCOT (measured by delivered spark spreads) being relatively tame until GSI surpasses 60% — and becoming extremely volatile as GSI rises above 70%.

To understand grid risks on spring days, we use Sunairio’s historical simulations in Figure 6 to plot expected (P50) and extreme (P99) hourly GSI in both early (March 15) and late (May 15) spring 2025. For clarity, we shade the background darker above GSI = 60% to reflect increasing price risk. As the plot shows, while the majority of hourly conditions in spring fall below the 60% level corresponding to high prices/capacity shortages, extremes in several hours present significant risks.

In particular, cold mornings and low renewables can drive hours ending (HE) 7–9 to the danger zone in early spring (left panel, Figure 6), while afternoon heat and decreasing solar in the evening drives HE 20–22 even further beyond in late spring (right panel, Figure 6).

Indeed, on the exceptionally cold early spring ERCOT morning this year (April 7), we saw prices spiking close to the price cap in ERCOT and averaging $1,552/MWh in an hour (Figure 7) due to high GSI resulting from exactly this combination of low temperatures (dipping into the 30ºFs surrounding Dallas, low renewables (minimal solar at HE7), and high generation outages (25+ GW).

That one hour on April 7 added approximately $4/MWh to the monthly peak price, meaning that just 0.2% of hours in the April peak contract (1/352) accounted for 10% of the entire contract value.

Looking ahead to 2026

With the spring drama almost behind us for 2025, what should we expect next year? Accounting for load growth and new unit additions/retirements, we find that conditions should generally be calmer next year across all spring hours. In other words, generation additions will outpace load growth — though the risk of acute price spikes remains, especially in the evening. As Figure 8 shows, P50 and P99 hourly GSI in both early spring (March 15) and late spring (May 15) are expected to be lower in 2026 compared to 2025, representing downside risk to ERCOT market prices.

Navigating spring volatility

Spring volatility highlights the challenges of managing grid and energy markets risk without a robust framework for evaluating the jointly correlated, multidimensional, granular, and skewed nature of the electric grids. To effectively navigate this period, grid planners and energy traders need to understand the complicated interplay between averages and extremes — to expect not just the seasonal trends but also to expect extreme events that drastically alter reliability and economic outcomes. It’s not enough to know that ERCOT grid balances for HE 7 typically don’t present much of a risk. We need to quantify the risk — via, for example, stochastic simulations — that weather and grid conditions conspire to drive capacity reserves low and market prices to the moon (as they did during HE 7 on April 7, 2025).

Introduction

“Prediction of instantaneous weather patterns at sufficiently long range is impossible.” This statement from Edward Lorenz remains as true today as when it was first written by the chaos theory founder in 1982, with current numerical weather prediction (NWP) model skill rapidly decreasing to zero over the course of a couple of weeks (Figure 1, also see Zhang 2018).

Although attempts to make deterministic predictions of localized and instantaneous weather devolve into chaos within a two week forecast horizon, it is possible to meaningfully predict large-scale atmospheric phenomena on seasonal time scales (several months to a year). These model predictions, however, need to be carefully processed before use–as the ECMWF states, “use of the raw numerical forecast products without interpretation is not recommended” (ECMWF 2021).

In this blog post, we (1) discuss one of the most impactful large-scale climate phenomenon (the El Niño - Southern Oscillation), (2) describe how Sunairio performs statistical processing and model assimilation to incorporate the latest El Niño intelligence into our climate simulations, and (3) evaluate how the resulting ENSO-informed climate simulations compare to both historical patterns and NOAA’s seasonal-timescale Climate Forecast System (CFS).

El Niño - Southern Oscillation

The El Niño - Southern Oscillation (ENSO) cycle refers to periodic changes in sea level air pressure and sea surface temperature (SST) in the southern Pacific Ocean. It is typically measured by the mean sea surface temperature anomaly of the “Niño 3.4” region shown in Figure 2, left panel (known as the Niño 3.4 index).

In the United States, especially high Niño 3.4 values (an “El Niño event”) are associated with warmer and drier winter weather in the northern US as well as cooler and wetter winter weather in the southern US (Figure 2, right panel, Halpert 2014). On the other hand, especially low Niño 3.4 values (a “La Niña event”) are associated with cooler and wetter winter weather in the northern US as well as warmer and drier weather in the southern US.

Figure 2: Regions of ENSO sea surface temperature indices (left) and the typical weather pattern expected in winter due to an El Nino event (right) (Source NOAA)

 

As the ENSO cycle is the “strongest interannual signal in the global climate system” (Tang 2018), it is heavily studied and forecasted, with current models showing positive predictive skill 6-12 months out (Barnston 2012). Figure 3 shows historical values of the Niño 3.4 index (left) and predictions from the CFS (right).

Figure 3: Historical Nino 3.4 region sea surface temperature anomaly (left) and current forecast of Nino 3.4 SST anomalies from the NOAA CFS (right)

 

Creating ENSO-aware Sunairio Simulations

Surveying the predictive power of current weather science, we find that high-resolution NWP models are predictive for at most 15 days, medium-resolution predictions of the ENSO cycle are predictive for 6-12 months, and low-resolution models of long-term climate trends can be predictive for longer time periods.

Sunairio’s climate simulations aim to therefore combine the best-available intelligence at each time horizon: long-term climate trends from the latest climate models (CMIP6), medium-term ENSO predictions from NOAA, and simulated hourly anomalies from Sunairio’s proprietary stochastic simulation generator.

Concretely:

1. Sunairio simulates 1,000 probabilistic hourly climate-trend aware weather paths from 15 days to 15 years (picking up where the NWP models lose skill) as correlated anomalies from climatological means that reflect historical weather patterns and adjust for CMIP6 climate trends.
2. Sunairio generates 1,000 probabilistic ENSO paths over a 12-month time horizon by extrapolating a range of outcomes from the latest 40 CFS model runs (4 runs per day times 10 days) (Figure 5, left panel).
3. Sunairio derives, from historical data, the correlation and effect of Niño 3.4 Index values on local weather variables throughout the year (Figure 5, middle panel).
4. Sunairio links hourly weather simulations (1) with ENSO paths (2)–and adjusts the simulated weather according to the corresponding Niño 3.4 index effect (3)(Figure 5, right panel).

 

Inspecting the January 2025 ENSO-informed Sunairio climate simulations in Figure 5, for example, we find that the median ENSO forecast is approximately -1σ (left panel), that the corresponding ENSO temperature effect at DFW is approximately +0.5^oC (middle panel), and that the ENSO adjusted Sunairio simulations at DFW are indeed approximately +0.5^oC warmer than January 2025 climatology (right panel).

Do Sunairio’s ENSO-adjusted Weather Simulations Reproduce Expected ENSO Effects Across Multiple Geographies and Weather Variables?

In this section, we verify that Sunairio simulations appropriately reproduce the historical relationship between the Niño 3.4 index over CONUS for major weather variables (2m temperature, 100m wind speed, and irradiance).

At 191 weather stations across the contiguous United States, we performed a backtest in which we simulated local hourly weather adjusted with historical Niño 3.4 Index values between the years 1997-2022.

Dividing calendar years into seasons (Dec - Feb, Mar - May, Jun - Aug, and Sep - Nov), we then calculated the impact of a +1σ Niño 3.4 value with monthly simulated weather averages at each site (Figure 7, right)–and compared these patterns to the historical (1950-2023) relationship between Niño 3.4 deviations and monthly weather averages (Figure 6, left). As we can see in Figure 6, Sunairio simulations accurately reproduce the historical ENSO relationship. (Click the weather variable headings to select different plots.)

Figure 6: Historical (left) and simulated (right) relationships between weather variables and Niño 3.4 index values. A positive +1σ Niño 3.4 index in winter, for example, tends to cause a 1 ^oC increase in monthly 2m temperature in Minnesota. Sunairio simulations faithfully reproduce historical relationships.

 

Predictive ENSO Signal + Stochastic Climate Simulation = High-fidelity Seasonal Forecast

As Sunairio simulations combine skillful ENSO predictions with high-fidelity climate simulations, medium-term Sunairio simulations are actually more predictive of future local weather than the CFS.

We validated this result by comparing historical monthly mean temperature measurements at the same 191 weather stations to both CFS predicted means and Sunairio simulated means. We found Sunairio simulations to be more predictive than CFS counterparts at 60% of the weather stations with an average skill score improvement of 2.1%. 

Incorporating high-fidelity climate modeling, in other words, makes Sunairio simulation averages more predictive than the CFS.

Conclusions

Reviewing the sections above, we find that:

• The El Niño – Southern Oscillation (ENSO) cycle is strongly correlated with global weather patterns.
• While deterministic models cannot predict instantaneous weather beyond a 15-day time frame, the Niño 3.4 Index can be forecast with some skill over a seasonal (6-12 month) time frame.
• Sunairio incorporates the latest intelligence on climate trends (CMIP6) and Niño 3.4 (CFS) predictions into its weather simulations.
• Sunairio simulations accurately reproduce the historical relationship between ENSO and local weather.
• Sunairio simulations, by combining high-fidelity climate simulation with a predictive ENSO signal, are more predictive of weather means than seasonal weather models.

Finally, we note two additional advantages of high-resolution Sunairio simulations over the CFS. First, Sunairio simulations can be generated at arbitrary locations, while CFS predictions must be interpolated from a 56km horizontal-resolution grid. Second, while seasonal forecast models like the CFS only offer one possible view of future weather in the medium term, Sunairio simulations offer 1000 paths of future weather for the next 15 years. As such, Sunairio weather simulations can be used to derive probabilistic risk distributions for commercial applications.

References

Barnston, Anthony G., Michael K. Tippett, Michelle L. L'Heureux, Shuhua Li, and David G. DeWitt. "Skill of Real-Time Seasonal ENSO Model Predictions during 2002–11: Is Our Capability Increasing?". Bulletin of the American Meteorological Society 93.5 (2012): 631-651.

ECMWF. “ECMWF System 4 user guide.” 1.1 (2011): 1-41. 

ECMWF “SEAS5 user guide.” 1.2 (2021): 1-44.

Halpert, M. “United States El Niño Impacts.” NOAA (2014).

Lorenz, Edward N. "Atmospheric predictability experiments with a large numerical model." Tellus 34.6 (1982): 505-513.

Saha, Suranjana, Shrinivas Moorthi, Xingren Wu, Jiande Wang, Sudhir Nadiga, Patrick Tripp, David Behringer, Yu-Tai Hou, Hui-ya Chuang, Mark Iredell, Michael Ek, Jesse Meng, Rongqian Yang, Malaquías Peña Mendez, Huug van den Dool, Qin Zhang, Wanqiu Wang, Mingyue Chen, and Emily Becker. "The NCEP Climate Forecast System Version 2". Journal of Climate 27.6 (2014): 2185-2208.

Tang, Youmin, Rong-Hua Zhang, Ting Liu, Wansuo Duan, Dejian Yang, Fei Zheng, Hongli Ren, Tao Lian, Chuan Gao, Dake Chen, Mu Mu. “Progress in ENSO prediction and predictability study.” National Science Review 5.6 (2018): 826–839.

Zhang, Fuqing, Y. Qiang Sun, Linus Magnusson, Roberto Buizza, Shian-Jiann Lin, Jan-Huey Chen, and Kerry Emanuel. "What Is the Predictability Limit of Midlatitude Weather?". Journal of the Atmospheric Sciences 76.4 (2019): 1077-1091.

Overview

What’s the best way to model power market dynamics that drive energy prices? In this blog post, we review traditional models via generation stacks, introduce the Grid Stress Index as a way to modernize traditional approaches for renewable growth and contemporary market realities, and finally show how well this model works out of sample by comparing summer 2024 electricity prices in Texas (ERCOT) to pre-season Sunairio distribution estimates.

The Traditional Approach: Grid Dispatch and the Generation Stack

Regional power grids are in balance between supply (generation and imports) and demand (customer load and exports). To minimize unnecessary costs, regional grid operators continually dispatch the least-cost mix of generation that meets demand while accounting for both physical constraints (e.g. transmission line capacity) and techno-economic constraints (e.g. minimum unit run times, minimum reserve requirements). 

Reflecting this least-cost mandate, regional energy prices are traditionally modeled via an idealized generation stack supply curve that assumes that all resources are dispatchable and ignores congestion and losses.

In Figure 1, for example, a demand of 4,000 MW would correspond to a least-cost generation mix that first dispatches all wind, solar, and hydrothermal generation and then dispatches a portion of combined-cycle gas generation. The marginal energy price would correspond to the marginal cost of combined-cycle gas generation ($25).¹

Unfortunately, not all grid capacity is available at any point in time. Power plants have outages and maintenance; wind and solar generation potential is weather-dependent. The widths of the blocks in Figure 1 (and thus the capacity vs price relationship) is constantly in flux.

A Modern Approach to Modeling the Power Grid Supply/Demand Balance via the Grid Stress Index

To modernize the traditional formulation of price as a function of demand, we first remove non-dispatchable renewables from the supply curve and instead net them against demand. The resulting measure of effective demand, net demand, is then the amount of energy that needs to be met with dispatchable capacity.

Net demand = Demand - (Non-dispatchable Renewables)

To normalize across periods with different demand levels and available capacity, we then transition from raw energy space to percentage space. A net demand of 7,000 MW when the grid has 10,000 MW of dispatchable capacity, for example, is modeled as 70% utilization of dispatchable resources.

Net Demand / Dispatchable Capacity

Next, we adjust the dispatchable capacity by the amount of generation outages to yield available dispatchable capacity, giving us

Net Demand / Available Dispatchable Capacity

or

Grid Stress Index = (Demand - (Non-dispatchable Renewables)) / (Dispatchable Capacity - Outages)

which is a single metric ranging from 0% to 100%, that describes the tightness of the power grid.

Finally, we map this metric to delivered spark spread prices instead of raw marginal prices, allowing us to normalize for the effect of changing fuel prices. We assume a standard heat rate of 6.5 MMBtu / MWh

Spark Spread = LMP - Heat Rate * Delivered Gas Price

The Relationship Between the Grid Stress Index and Market Prices

Moving from theory to practice, Figure 2 plots our Grid Stress Index metric against delivered spark spreads for all hours June-August 2023. 

The relationship looks very similar to the idealized marginal supply curve (Figure 1). We can clearly see a slow rise in marginal cost followed by an inflection point around 75% available dispatchable capacity utilization, after which prices rise swiftly. Interestingly, the extremes of this plot are higher than those modeled by the traditional approach, as prices are often well above physical fuel costs (reflecting scarcity-seeking behavior from market participants). Putting this together, we’ve effectively learned the true ERCOT marginal supply curve without actually modeling the individual unit generation stack. 

Using Sunairio Simulations to Model Grid Stress Index 

Now that we’ve established the fundamental relationship between our measure of the Grid Stress Index and energy prices, we note that realistic simulation of this measure requires correctly producing inputs that are probabilistic (replicating extreme events at the correct likelihood) and fully correlated (replicating the correct joint distribution of load, renewables, and generation outages). The Sunairio simulation suite is purpose-built for this task.

In order to present an up-to-date view of energy resources, we update installed capacity projections monthly and retrain/re-simulate energy models weekly. 

As we can see from Table 1, there are significant trends in both installed generation capacity and demand, making the downstream effect on the hourly Grid Stress Index complicated to anticipate.

Installed Capacity	2023 Jun-Aug	2024 Jun-Aug	Change
Wind	37.7 GW	39.6 GW	+1.9 GW
Solar	16.6 GW	27.1 GW	+10.5 GW
Non-renewable	89.2 GW	90.3 GW	+1.1 GW
Battery	2.6 GW	7.2 GW	+4.6 GW
Average Load Growth			+7%

What Did Sunairio Say About ERCOT Summer 2024 Grid Risks?

To benchmark Sunairio ERCOT simulations against historical realizations, we compared our pre-summer simulated distribution of Grid Stress Index to the actual distribution from June through August 2024. ²

Figure 3 (left) presents the distribution of forward-looking Sunairio simulations from May 3, 2024 (blue line) against actual distribution (green histogram). The fit is quite close, implying that Sunairio simulations from early May accurately reflected the various risks to the hourly Grid Stress Index (alternatively, one can interpret this fit as an indication that ERCOT June-August 2024 hourly grid balances were distributed close to expectations).

However, comparing the 2023 and 2024 historical distributions of the Grid Stress Index (Figure 3, right), we see that the ERCOT market was much tighter in 2023 than in 2024. In particular, 2023, had far fewer hours in the high-priced regime above 75% available dispatchable generation utilization–a shift that we see ultimately reflected in realized prices too (Figure 4): 

North Hub Delivered Spark Spread (6.5HR)	2023 Jun-Aug Actual	2024 Jun-Aug 5/3/24 Market	2024 Jun-Aug Actual	2024 Actual vs 5/3/24 Market	2024 Actual vs 2023 Actual
Peak ($/MWh)	$153.63	$112.85	$29.00	-$83.85	-$124.63
Off-peak ($/MWh)	$27.09	$45.06	$14.95	-$30.11	-$12.13
7x24 ($/MWh)	$86.69	$76.79	$21.47	-$55.32	-$65.22

Looking at Table 2, we note that 2024 realized prices were not only much lower than their 2023 counterparts but also much lower than pre-season market forward prices.  Sunairio pre-season ERCOT simulations at that time, however, were much closer. In other words, Sunairio simulations are excellent predictors of 1-3 month-forward grid balances and are also effective market signals of significant downside to energy prices relative to the same period from 2023.

Conclusions

1. The Grid Stress Index is an improved metric for modeling power grid supply-demand balances that inherently accounts for renewables variability and unit outages.
   
2. The hourly distribution of the Grid Stress Index showed that the June-August 2024 period in ERCOT had significantly less risk of high-priced hours compared to the same period in 2023.
   
3. Sunairio simulations of summer ERCOT grid balances–generated in May 2024–were an extremely accurate forecast of actual grid conditions.
   
4. Sunairio simulations from May 2024 were an accurate signal of downside price risk in ERCOT for the June-August 2024 period relative to the same period from 2023. 

---

1. Battery storage will also appear in the supply stack at varying marginal costs depending on their cost of charging and their efficiency. ↩︎
2. The conventional power market definition of summer is limited to July-August, though we add June as well in this analysis to increase sample size. ↩︎

Introduction

Any business that makes money from the ERCOT power market knows that not all price hours are equally important. Over the past 5 years, for example, the top 1% of price hours have accounted for an average of 34.4% of the annual spot energy market value (Table 1).

	Share of Annual Market Value from Top 1% Price Hours, ERCOT North Hub RT
2019	35%
2020	17%
2021	60%
2022	20%
2023	40%

Energy prices are heavily skewed and volatile, driven by net load volatility resulting from combinations of high heating/cooling demand and simultaneous low renewables generation (Figure 1). Properly accounting for extremes, in other words, is essential for survival in a rapidly changing market.

A primary driver of the changing power market landscape, of course, is climate change. Unfortunately, most power market participants rely on unadjusted historical weather data to estimate future market risk (a practice founded on the hidden assumption that the future is like the past) and don’t account for changes in underlying weather distributions over time.

In this blog post, we demonstrate the outsized effect that a shift in weather distributions due to climate change has on the frequency of extremes in both temperature and total system load.

Climate Change and the Expectation of Extremes

Climate change trends are usually studied and reported as expected differences in average magnitudes over a period of time. For example, analysis of historical temperatures across Texas yields a trend of approximately +0.5°F per decade over all of Texas, consistent with CMIP6 climate models. In reality, however, these trends are better understood as shifting entire temperature probability distributions to the right–and not just the resulting shifts in aggregate measures like means and medians (Figure 2). As the probability of an extreme can be understood as the area under the distribution to the right of a particular point, we see from Figure 2 that a shift in distributions has a particularly dramatic effect in the relative frequency of extremes.

Changing Extreme Expectations in ERCOT

To illustrate the outsized effect of climate change on the frequency extreme occurrences of temperature and load in ERCOT, we simulated 1,000 probabilistic hourly weather outcomes for 2024 under two conditions: 1) using unadjusted historical weather data from 1950-present, and 2) applying climate change trends to adjust historical data to current climate conditions. Figure 3 presents the increased occurrence of temperature extremes (population-weighted 100°F+ hours) and load extremes (80 GW+ of RTO load) during all summer hours when accounting for climate change.

The temperature heatmap (top) is fairly straightforward–the effect of increased extreme frequency due to climate change is greatest in early August (when ERCOT temperatures are highest) and centered around the hours beginning 15, 16, and 17 (local time). After accounting for climate trends, 100°F+ hours occur 80 more times over 1000 simulations in the early August late afternoon period, which equates to an 8% increase in the overall likelihood in each hour. 

The load heatmap is more interesting. First, one notes a semi-regular pattern due to decreased load during weekends and holidays. More importantly, however, one finds that the effect of climate change on the frequency of extremes is shifted by about a week or two. Even though extreme temperatures might be more frequent around August 1, extreme load is more frequent around August 7–due to the natural seasonality of cooling demand intensity in Texas. Finally, the heatmap values are significantly different. Accounting for climate trends leads to an additional 300 occurrences of 80 GW load hours during the early August, late afternoon period, or a 30% increase in overall likelihood in those hours.

Turning to the more relevant market risk metric of relative extreme event frequency increases, we compare the outright likelihood of events in the climate-aware simulations and the unadjusted simulations. Figure 4 presents the outright likelihoods of 80+ GW load in hour beginning 19 in terms of their occurrences per 1,000 climate simulations. In August, extreme load occurs in approximately 5% of simulations that don’t incorporate climate trends, but almost 25% of simulations that do. That’s an over 6x increase in odds of having extreme (80 GW) load (1/19 vs 1/3).  

Conclusions

• Climate change alters mean weather expectations slightly but causes significant increases in the frequency of extreme weather.
  
• Weather-affected businesses experience disproportionately large profit fluctuations as the occurrence of extreme weather events increases.
  
• In the ERCOT power market, prices follow an extremely skewed distribution, with only 1% of hours per year accounting for more than 30% of total value.
  
• Traditional approaches for estimating the likelihood of extreme events in power markets rely on historical data and thus suffer from changes in non-stationary weather time series.
  
• When properly accounting for climate change in a current-year weather simulation, the frequency (in absolute terms) of extreme temperature hours (100°F) increases up to 8%; the frequency of extreme ERCOT load hours (80 GW) increases up to 30%.
  
• In relative terms, the frequency of extreme load (80 GW) hours increases by up to 6x for the 19:00 hour.

Stay tuned for future discussion of the climate change effects on net load!

Introduction

During a major volcanic eruption, powerful volcanoes may launch sulfur gas high into the stratosphere, where it can form long-lasting aerosols persisting on time scales from months to years. These clouds of aerosols then diffuse into a layer that can cover the globe. Studies have shown that this process can result in a severe reduction of surface solar irradiance and reduce clear sky irradiance by as much as 13 percent (Robock 2000), with significant implications for solar PV production [Robock 2000, Bluth 1992, Hay and Darby 1984]. 

Obviously, the threat of these months to years-long volcano-induced periods of low solar irradiance presents a major risk to solar investor profitability and regional power grid reliability. However, these risks aren’t typically considered when investors and planners calculate expected production variability from new (or existing) utility-scale solar sites–primarily because conventional risk assessment approaches rely on a limited sample of historical irradiance data that coincidentally corresponds to a time period (beginning in 1998) without any of the major volcanic events that have affected North America.

A Limited Historical Record of Irradiance Data 

Due to the dearth of high-quality ground station solar irradiation measurements, the solar industry has come to rely on remote sensing data from geostationary satellite missions. Irradiance data from geostationary satellites is available starting in 1998. The resulting tradeoff between data quality and data breadth caused by this exclusive reliance on satellite data for production risk assessment evinces itself as a major gap in understanding just how low solar production can be if and when the next planetary scale volcanic eruption occurs. 

To put volcanic eruption risks to solar production into context, we estimate (as shown below) that volcanic aerosols can produce annual solar irradiance deficits 5 standard deviations below the satellite record estimate. In other words, using data from 1998-2023 to estimate an annual solar irradiance probability distribution at a solar site (as is commonly done today), a major volcanic eruption could lead to a year of observed solar irradiance 5 standard deviations lower than the mean expected value. Naïvely, a 5-sigma deviation is extremely unlikely: for example, if we assumed that annual irradiance was normally distributed, a 5-sigma event should happen on average 1 in 3.5 million years.

Yet we know that these risks are much more common–two eruption events occurred between 1982 and 1991–meaning that deriving a measure of irradiance variability using only the 1998-forward satellite record probably underestimates volcano risk. The industry, in effect, may be prioritizing short timescale accuracy at the expense of long-term risk assessment.

El Chichón and Pinatubo

The last fifty years have seen two eruptions with large scale persistent impacts to irradiance. The eruption of El Chichón in 1982 accelerated roughly 7 million tons of sulfur dioxide to atmospheric heights of over 22 kilometers. The eruption of Mt. Pinatubo a mere 9 years later resulted in a 35 km high plume of 20 million tons of sulfur dioxide (Bluth et al. 1992). Both plumes became entrained in the atmosphere, reacted with water vapor, and spread out to form reflective bands of sulphuric acid aerosols that encircled the globe within roughly three weeks and persisted for up to 3 years after their respective eruptions.¹

Although satellite data is not available before 1998, the Mauna Loa observatory in Hawaii captured the impact of these global aerosol blankets on solar irradiance through measurements of the apparent transmission of the atmosphere during hours with clear skies. The data, presented in Figure 1, shows clear sky irradiance dropping by more than 13% due to the impact of El Chichón in 1982 and 11% due to the impact of  Pinatubo in 1991.

To understand how these reductions in irradiance impacted the solar resource in the US, we aggregated surface solar irradiance from the ERA5² dataset in the southwestern states (CA, UT, CO, NV, AZ, NM) from 1950-2023. We found that the 4 years with the lowest mean irradiance on record followed these eruptions. When compared to the period since 1998, the year following El Chichón had an irradiance deviation of -5.3 𝝈; the year following Pinatubo had an irradiance deviation of -3.8 𝝈.

10 Lowest Surface Solar Irradiance Years: ERA5 1950-2023 (Avg of CA, UT, CO, NV, AZ, NM)
Rank	Year	Mean Annual GHI (W/m2)	Mean Annual GHI 1998-2023 (W/m2)	𝝈1998-2023	Notes
1	1983	221.5	234.9	-5.3	Year after El Chichón
2	1982	225.4	234.9	-3.8	El Chichón
3	1992	225.4	234.9	-3.8	Year after Pinatubo
4	1984	227.0	234.9	-3.3	2 years after El Chichón
5	1986	227.7	234.9	-2.9
6	1981	228.6	234.9	-2.5
7	1957	229.0	234.9	-2.3
8	1991	229.6	234.9	-2.1	Pinatubo
9	1987	229.9	234.9	-2.0
10	1998	230.3	234.9	-1.8

To visualize these years within the full ERA5 date range, Figure 2 plots the time series of regional annual average solar irradiance from 1950-2023. Again, we see that the years following El Chichón and Pinatubo are the lowest, appearing as outliers compared to the 1998-forward satellite record range (consistent with the 3+ sigma measurements above).

Data Challenges and Future Risk

As seen above, volcanic eruptions can be an enduring hazard to solar production in the United States. However, because historic eruptions are not captured by the satellite record, they are often not accounted for during due diligence or planning exercises. Moreover, as the solar PV industry has not existed at scale for long enough to have meaningful first hand experience with volcanic eruption related deficits, empirical or even anecdotal evidence of their effects is rare.   

Clearly, solar production risk assessments using only historical data since 1998 are not incorporating these relatively rare but impactful events. Yet the risk of the next major eruption over the next decade may be greater than the industry realizes. Estimates of major eruption frequency (Rougier et al. 2018, Sheldrake and Caricchi 2017, Pyle 1995), though imprecise, suggest that volcanoes at least as severe as Pinatubo and El Chichón may occur with a frequency of roughly 1 in 51 years (Sheldrake and Caricchi 2017)–corresponding to a 33% chance of occurrence during a 20 year solar project life cycle.

Sunairio Irradiance Simulations Reveal Long Tails Consistent with Volcanic Events

Overcoming the limitations of a small sample set when estimating risk is a key benefit of employing stochastic simulation methods. While grounded in high-quality satellite-based irradiance data, Sunairio’s stochastic climate simulations generate a wide range of realistic weather scenarios that can extrapolate beyond the limitations of a restricted date range, using statistical and machine learning methods to infer complex spatio-temporal relationships between weather variables, locations, and timepoints to create a broad data set from which to evaluate risk. Therefore, Sunairio’s simulations are both representative of local weather down to a few kilometers and useful for characterizing variability–including extreme events.

To demonstrate how Sunairio simulations can “see” weather risk beyond a limited historical record, we simulated 1,000 scenarios of hourly weather for 2024 at a location just outside of Albuquerque, New Mexico–using satellite-based irradiance data from 1998-2023. Figure 3 plots the annual average GHI of each of our simulations compared to two distributions from ERA5: the 1998-2023 period (satellite data period), and the full 1950-2023 period of record. 

In the left plot of Figure 3 we see that the Sunairio simulations show a left skew (risk of low solar irradiance) that notably isn’t present in the 1998-2023 ERA5 distribution–but is present in the right plot of historical data since 1950, overlapping with the occurrence of major volcano years. Sunairio simulations, in other words, replicate the true tail of the historic distribution (including years with volcanic activity) which isn’t seen in the limited satellite data period.

Conclusions

In this case study we showed that:

1. Major volcanic eruptions such as El Chichón (1982) and Pinatubo (1991) can significantly reduce surface solar radiation by creating a planetary-scale band of sulphuric acid aerosols that persist in the atmosphere for years. The effect is well documented by Moana Loa station observations.
   
   
2. Traditional solar PV production risk estimates do not incorporate the risk of volcanic events because they derive annual risk estimates using satellite-based irradiance data (which starts in 1998).
   
   
3. The signatures of El Chichón and Pinatubo eruptions are apparent in ERA5 reanalysis data: the four lowest mean annual irradiance years (between 1950 and 2023) are either eruption years or years immediately following those eruptions.
   
   
4. Using the 1998-2023 satellite data period as the basis for risk estimates, volcanic eruption years appear to be 3- to 5-sigma events (i.e. extraordinarily unlikely)–contradicting the research on volcanic eruption frequency.
   
   
5. Sunairio’s simulations replicate low-irradiance years at a frequency consistent with the historic record and reanalysis estimates. Using Sunairio simulations of Albuquerque, NM for 2024, the likelihood of a year having mean annual GHI equivalent to the 1983 El Chichón year is roughly 2.3%.

Notes

1. Not all eruptions disrupt the climate on a planetary scale. The eruption of Mt. St. Helens in 1980 and, more recently, the eruption of Hunga Tonga–Hunga Haʻapai in 2022 both produced large explosions but failed to propel large quantities of SO₂ into the stratosphere. Therefore, although they had catastrophic impacts on surrounding areas, the plume of aerosols were able to settle to the earth within a few weeks. ↩︎
2. Although ERA5 data is less accurate than satellite data, these errors are most extreme during cloudy conditions (Urraca 2018). Looking at states in the southwest (which have fewer cloudy hours) allows us to partially mitigate these errors. ↩︎

Citations

Bluth, G. J. S., et al. (1992) Global tracking of the SO2 clouds from the June 1991 Mount Pinatubo eruptions, Geophys. Res. Lett.

Hay, J. E. and Darby, R. (1984) El Chichón – influence on aerosol optical depth and direct, diffuse and total solar irradiances at Vancouver, B.C., Atmosphere-Ocean, 22:3, 354-368, DOI: 10.1080/07055900.1984.9649204

Pyle, D. M. (1995) Mass and Energy Budgets of explosive volcanic eruptions. Geophysical Research Letters; 22, 5 563-566.

Robock, A. (2000) Volcanic Eruptions and Climate. Reviews of Geophysics.

Rougier, J., et al. (2017) The global magnitude–frequency relationship for large explosive volcanic eruptions. Earth Planet. Sci. Lett.

Sheldrake T., and Caricchi L. (2017) Regional variability in the frequency and magnitude of large explosive volcanic eruptions. Geology; 45 (2): 111–114. doi: https://doi.org/10.1130/G38372.1

Urraca, R. et al. (2018) Evaluation of global horizontal irradiance estimates from ERA5 and COSMO-REA6 reanalyses using ground and satellite-based data. Solar Energy.