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.

Sunairio is excited to announce it has been awarded a National Science Foundation (NSF) Small Business Innovation Research (SBIR) Phase I grant. The NSF SBIR program, which selects less than 14% of applying startups, funds high-impact technological research and development (R&D) across a wide range of disciplines.   

For Sunairio, the award will fund research into a new, highly scalable climate simulation engine that promises to generate climate insights for the energy sector at 1,000 times the resolution of traditional, physics-based global climate models, such as the ones which support the body of research conducted by the Intergovernmental Panel on Climate Change (IPCC). 

Sunairio’s new climate simulation solution is a novel combination of generative AI and advanced statistics that early research shows can replicate forward-looking local (3-km x 3-km) hourly weather patterns and climate trends for up to 15 years. This approach will fill a technical and commercial gap because the traditional global climate models are very computationally intense and, therefore, only run at 100-km x 100-km, which isn’t sufficient to accurately model site-specific weather risk at wind and solar farms. 

Read the full press release here.

Sunairio co-hosted a webinar with Yes Energy that demonstrated how to build actionable price models for term power markets that allow you to generate price distributions which incorporate weather volatility and renewables capacity scenarios.

REQUEST ACCESS TO THE WEBINAR

With ERCOT reporting almost one-third (30GW) of non-renewable generation capacity offline as of November 12th, it’s clear that we’re in the height of fall outage season. Accordingly, let’s take a look at generation outage trends, patterns, and the risk that extreme cold this winter will force generators offline.

For this blog post we’re going to focus on outages of non-renewable generation capacity, which could also be thought of as controllable–or thermal–generation capacity. In other words, we’re not focusing on outages of wind and solar resources (which we’ll tackle in a separate piece).

The Seasonality and Variability of Generation Outages

Figure 1 plots maximum daily RTO load and average daily generation outages in ERCOT for each year since 2017.  Both are seasonal, with generation outages increasing during the Spring and Fall shoulder periods of lowest load. Spring generation outages tend to be higher than Fall outages, even though spring loads are slightly higher–-likely because generation owners want to ensure that maintenance and repairs are completed before the high-demand (and potentially high-priced) Summer season. 

Variability around the seasonal trend is also very apparent, with Spring and Fall outages varying between 20 and 30 GW and even Summer outages exhibiting a range (from the lowest observed outage levels to the highest observed levels) of at least 5 GW–more than enough to tip ERCOT pricing from moderate to extreme levels in a typical Summer season.

Beyond the normal variability, a couple of generation outage events stand out as outliers–in particular February 2021 (light purple) and August/September 2017 (dark purple). As winter is coming, let’s take a closer look at the February 2021 event and see if we can infer the likelihood of a similar event occurring again over the next few months. 

How Cold Does it Need to Get in Texas for Outages to Spike in Winter?

In short: quite cold for Texas. Excess February 2021 generation outages were caused by Winter Storm Uri, which in turn knocked out power for much of the state over a week-long period. Figure 2 plots the daily minimum ERCOT temperature (population-weighted) against average daily outages for December, January, and February.

We see essentially no correlation between temperature and generation outage levels–with the notable exception of extreme cold starting on February 15th, 2021 forcing approximately 10-15 GW of generation offline  (with respect to February 13th and 14th–right before the cold snap). 

Outages climbed significantly when temperatures bottomed out close to 10 degrees F and stayed elevated even after the daily lows increased, leading one to suspect the existence of a cold temperature threshold that’s likely to cause excess generation outages that may persist for days (or longer). Looking at February 2021, we estimate that this threshold is somewhere around 10 degrees F. 

Of course, while this analysis evaluates weather conditions across the entire ISO in order to infer an estimate of ISO-wide excess generation outages, the weather conditions that force generators offline are local. To make more precise cold-weather outage models, we’d want to repeat this analysis by grouping generators according to regional weather zones and then comparing local weather to regional outages.    

Will ERCOT Rule Changes Prevent Excess Cold-weather Outages From Happening Again?

We’re skeptical. The primary rule change that ERCOT instituted (at the behest of the Public Utility Commission) is a weatherization standard that will be “enforced”-without penalties. To be in compliance with the standard, generators are to:

1. Implement additional weather emergency preparation measures reasonably expected to ensure sustained operation at 95th percentile minimum average 72-hour wind chill value in the ERCOT weather study for facility’s weather zone
   
2. Create list of all cold weather critical components, review annually prior to beginning of winter, and update as necessary

(Source)

In other words, there’s still no economic penalty for an ERCOT generation resource that’s unavailable in a grid emergency. As a consequence, most thermal generators (e.g. natural gas plants) still do not have firm fuel supply arrangements with either pipelines or commodity suppliers. That means that on very cold winter days, a natural gas generator is not only subject to outages resulting from its own mechanical failures--but also from failures of its natural gas supply (pipeline capacity is maxed out, lower daily gas production due to natural gas wells freezing). 

How Can Sunairio Quantify the Likelihood of Extreme-Cold-Induced Outages this Winter?

We can use Sunairio weather simulations to find out the odds of breaching the 10 F daily low temperature threshold we identified above. We execute the following query to get simulations from Snowflake:

select
    convert_timezone('Etc/UTC', 'US/Central', sim_datetime) as sim_datetime,
    sim_number,
    sim_value*1.8 + 32 as sim_value -- Converting C to F
from
    ercot_weather_sims
where
    location = 'rto'
    and variable = 'temp_2m'
    and convert_timezone('Etc/UTC', 'US/Central', sim_datetime) 
        between '2023-12-01 00:00' and '2024-02-29 23:59'

By downloading this with the Snowflake connector in Python, you can make the Pandas DataFrame a timeseries of simulations with a simple pivot:

cur = conn.cursor()
cur.execute(query)
temp_sims_df = cur.fetch_pandas_all().pivot(index='SIM_DATETIME', columns='SIM_NUMBER', values='SIM_VALUE')
cur.close()

Then, with the resulting 1,000 hourly simulations of ERCOT population weighted hourly temperature for Dec 1, 2023 through Feb 29, 2024 we can calculate the probability distribution of minimum daily low temperatures for each simulation path (in python, for example):

temp_sims_df.resample('d').min().min().hist()

This gives us the histogram in Figure 3. The height of the bars in this histogram represent the number of paths in our 1,000-path simulation with the corresponding daily minimum temperature on the the x-axis. Most of the distribution is around 22 degrees. In other words, we’d expect the coldest daily low temperature this winter to be about 22 degrees.

So what’s the likelihood of having a daily low temperature 10 degrees or lower? We just count the number of paths < 10 and divide by the total number of paths (1,000).

In  [1]: (temp_sims_df.resample('d').min().min()<10).sum() / 1000
Out [1]: 0.055

We find a 5.5% chance (according to these assumptions) of experiencing extreme cold that could result in excessive non-renewable generation outages this winter. 

Sunairio Outage Simulations Natively Account for Cold-Weather Effects

At first glance, excess generation outages similar to those caused by Winter Storm Uri may seem like outliers. However, armed with a theory (cold temperatures beyond a certain threshold cause a cascade of multi-day gas supply and gas generation problems) and forward looking simulation data (1000 weather simulation paths) we can model the chance and effect of another similar event.

At Sunairio, we model generation outages (including the effect of extreme cold weather) as a percentage of thermal energy in parallel with weather simulations-ensuring that data and analyses are not left in the dark.

Conclusions

• Non-renewable generation outages in ERCOT follow a seasonal pattern and peak in Spring, though with significant variability.
  
• On an ISO population-weighted basis, daily low temperatures need to drop to about 10 degrees Fahrenheit to cause excess generation outages.
  
• New ERCOT rule changes require power plants to attest to winter weatherization but still don’t provide a strict penalty for noncompliance or unavailability.

• The odds of a cold-weather-induced generation outage event happening during Winter 2023-2024 is about 1 in 20.
  
• Sunairio non-renewable generation outage simulations for ERCOT incorporate these dynamics.

Overview

Average 100m wind speeds in parts of the U.S. Midwest were lower than ever during May through July of this year (based on ERA5 data from 1950 to the present). For example, the plot below displays 30-day average 100m wind speeds for Clay County, MO (just outside of Kansas City) with 2023 highlighted in yellow:

Besides causing significant wind energy production and revenue shortfalls, this record low-wind period has precipitated questions throughout the wind energy industry related to the role that climate change is playing, the likelihood of similar events happening in the future, and strategies to mitigate impact.

Let’s start by looking at the trends themselves.

Annual and Seasonal Trends in 100m Wind Speed

On an annual basis, average 100m wind speeds have been declining in much of the Eastern and Midwest U.S., including wind-rich onshore areas of IL, IA, KS, NE, and MN as well as wind-rich offshore areas off the coasts of NJ, NY, CT, and MA. The following map shows wind trends across the US, with regions of statistically significant (p < 0.01) trends bounded by a black border.

Annual trends, however, don’t reveal differences throughout the year. For example, here’s a map of wind trends for January and August:

Note the strong negative trends in August throughout the northern Midwest and East Coast (including offshore) which are either not present or much less pronounced in January. In these areas, in other words, not only is It getting less windy—it is getting significantly less windy in summer.

What’s causing this?

Climate Change Impacts

At Sunairio, we get versions of this question all the time. “Is this climate change?”, “Is this going to get worse?”

What we know is that global winds experienced a pronounced decline starting around the 1980s. This “global stilling” has been less intense since 2010, though wind speeds are still below pre-1979 averages. The Intergovernmental Panel on Climate Change forecasts slowing wind speeds through 2100–at which point they predict that annual average wind speeds could drop by 10%. Unfortunately, compared to the body of knowledge on global temperature trends, there’s not as much scientific consensus on the exact causes of trends in wind speed––though two major drivers are thought to be controlling: 1) changes in large-scale atmospheric circulation and 2) an increase in “surface roughness” (e.g. forest growth, urbanization).

To visualize what’s happening more precisely, we created an animation of the probability distribution of summer wind speeds for a county in Kansas relative to 1950-1979.

This animation helps convey an important point about the relationship between climate trends–which are usually described by changes in averages over time–and the likelihood of extremes: a small shift in the center of a weather probability distribution may cause a large change in extreme event frequency. As the animation progresses, note that the likelihood of extremely calm (low-wind) periods is much more pronounced than the relatively small shift left of the overall wind speed distribution.

The Difference in Wind Power Production

With some wind power curve assumptions we can calculate how much less wind power was produced than expected during May-July 2023. As we see in the figure below, many areas with high concentrations of wind farms saw production deficits of 20%-30% translating to approximately 17,000,000 MWh less wind energy production than expected throughout the U.S.

Note that this is 20%-30% lower than expected wind power production estimates which already account for the observable, negative trends in wind speed over time (meaning that current-year expectations should generally be lower than historical averages). If we calculated the drop in wind power against a simple historical average, the deficit would be even greater.  

Financial Budgeting Implications  

Wind power production expectations are typically the most critical input in financial budgets for utility-scale wind farm owners. Traditionally, renewable energy asset owners have set production and revenue budgets at what they consider to be median (or P50) levels based on a backward-looking analysis of historical weather. 

In this context, an event such as the May-June 2023 period would be considered a loss (relative to budget)–and because the P50 expectations are usually not adjusted for climate trends, those losses would likely have been greater than the 20%-30% we estimated.

But even with budgets that account for declining wind speeds, asset owners face a second problem: how to anticipate production variability–i.e., the risk of a low wind or high wind periods. This variability is essential to estimate in order to ensure sufficient capital reserves to survive adverse events and anticipate the magnitude of cost-saving plans that might be necessary to hit annual profit targets. Traditionally, these variability assessments (also known as production risk assessments) have also been accomplished using history. As the May-July 2023 event shows, however, using a limited sample of historical wind speeds as a proxy for future wind speed risk–and not accounting for climate trends–does not work.

Grid Reliability Implications

Wind farm owners aren’t the only stakeholders who depend on wind power. Regional power grids increasingly rely on renewables for essential grid-balancing energy. Events such as the May-July 2023 low-wind period expose utilities and ISOs to significant risks because resource adequacy modeling uses historical weather data only, and this event was (by virtue of being a record) outside the modeled range of possibilities. 

Luckily, regional power grids did not suffer reliability events during this period, likely due to a range of factors–the most significant being an additional 12.5 GW of solar capacity being installed year over year (including almost 4 GW in what the EIA classifies as the “West North Central”) and a relatively mild summer in the upper Midwest.

What Does Work: Stochastic Climate Simulation

For all the issues described above, Sunairio incorporates forward-looking climate modeling to assess renewable energy production risks. Our solution is stochastic in nature, easily scalable , and done at very high temporal (hourly) and spatial (3km or less) resolutions. When we simulate 1000 outcomes of a future time period, it’s like having access to 1000 probabilistic climate-aware possibilities. 

With 1000 probabilistic outcomes of current and future-year wind speed, we can accurately quantify expected (or median/P50) wind production, and risks to that production, much more accurately (at least 4x more accurately) than is possible using historical data alone. Moreover, we can predict the likelihood of extreme weather events happening–even those that are statistically possible but not in the historical record (yet).

For example, our simulations show that the May-July 2023 event should occur at a frequency of about 1 in 83 years. Perhaps even more relevant: any 3-month period (not just May-July) with an equivalent wind production capacity deficit will occur roughly 1 in every 10 years. These are extreme events–but they will occur again, and they shouldn’t be considered unknowable risks.    

Conclusion

North American wind speeds are declining in some of the most wind-rich areas–where a significant amount of wind energy capacity is either already or soon to be installed. 

When future wind speed expectations are tied to historical averages and historical ranges as part of long-range planning exercises, these wind energy production estimates will eventually be wrong. 

A robust solution to this problem is to use forward-looking climate simulation for wind energy production risk assessment.

We’re excited to announce that Professor Matthew Lackner has joined Sunairio as an Advisor. Lackner is a Professor of Mechanical Engineering and the Director of the Wind Energy Center at UMASS Amherst. Prof. Lackner will be advising Sunairio on all areas of wind energy simulation.

Sunairio is proud to join EPRI's Climate READi Affinity Group. Climate READi is the industry's collaborative approach to managing climate risk in the power sector. Sunairio will be supporting the effort by providing our technical expertise in predictive climate and energy analytics.