Profitability and investment risk of Texan power system winterization

A lack of winterization of power system infrastructure resulted in substantial rolling blackouts in Texas in 2021, but debate about the cost of winterization continues. Here we assess if incentives for winterization on the energy-only market are sufficient. We combine power demand estimates with estimates of power plant outages to derive power deficits and scarcity prices. Expected profits from winterization of a large share of existing capacity are positive. However, investment risk is high due to the low frequency of freeze events, potentially explaining the under-investment, as do the high discount rates and uncertainty about power generation failure under cold temperatures. As the social cost of power deficits is one to two orders of magnitude higher than the winterization cost, regulatory enforcement of winterization is welfare enhancing. Current legislation can be improved by emphasizing the winterization of gas power plants and infrastructure. There has been continued debate on the cost and benefits of winterization of energy infrastructure in Texas, especially after 2021 power outages. Here the authors estimate the expected revenues from winterization, and its cost, and find that winterization would pay off taking into account the probability of similar outages.

W eather extremes such as storms can strongly affect the reliability of power systems 1 . The increasing use of variable renewable energies additionally exposes power systems to hazards caused by weather extremes 2,3 . However, recently a gas-power-dominated system was deeply impaired by a weather extreme: a cold spell over Texas between 10 February and 20 February 2021 with temperatures far below 0 °C caused a failure of large parts of the Texan power system. The combination of extraordinarily high winter electricity demand and more importantly the failure of substantial power generation capacities, both due to low temperatures, resulted in up to 4.5 million Texans being cut off from their electricity supply 4 .
A first retrospective analysis of the event by Busby et al. 5 discusses the magnitude of the event and its causes, indicating that the total economic loss amounted to US$130 billion and that the outage of gas power plants was mainly responsible for the high deficits in power generation capacity. Wu et al. 6 provide a power grid simulation to conduct a very detailed analysis of the 2021 event, and Doss-Gollin et al. 7 have shown that lower temperatures than those in February 2021 have been observed in the past 71 years, and heating demand predicted from temperature data would also have been higher in the past, although the 2021 freeze event was comparably long. These previous studies indicate a striking gap between the occurrence probability of such an event, its large-scale economic and social cost and the lack of winterization efforts. However, none of these studies assessed whether the economic incentives for power companies to invest in winterization have been sufficient, when the 2021 event is put into a long-term climatic context. As winterization was not strongly enforced by regulation in Texas, power generators had to rely on the incentives provided by the energy-only market to arrive at investment decisions. These incentives consist mainly of regulated price spikes at the spot market when generation capacity is scarce 8 .
Here we assess how revenues from winterization compare to its cost for power companies. Technically, we combine estimates of temperature-dependent load with a model of power plant outages, taking into account 71 years (1950-2021) of past climate from reanalysis data. Subsequently, we discuss the 2021 event in detail, analyse the long-term frequency of such events and determine revenues from and cost of winterization. Furthermore, we discuss potential reasons for under-investment.

The 2021 event in a long-term climatic context
Starting on 10 February 2021, temperatures in Texas decreased, causing load to increase from around 40 GW to over 70 GW by [14][15] February. On 15 February, the freeze reached a critical level and, consequently, substantial shares of generation capacities failed. Available capacities dropped below demand, leading to a sustained deficit in power generation capacity (Fig. 1). Consequently, rolling blackouts had to be implemented to stabilize the grid, and scarcity prices at the power market increased to the upper limit of US$9,000 MWh -1 . The deficit event continued until 19 February, when rising temperatures allowed the system to recover.
In the following, we compare events in Februrary 2021 to the period 2004-2020, as the Electric Reliability Council of Texas (ERCOT) provides hourly data on system operation in this period. Furthermore, the values of loss of load, capacity failures and demand prediction in this section rely on our simulation and may, therefore, differ from ERCOT reports to some extent. Because we focus on estimating the long-term frequency of such events, we did not aim to reproduce the February 2021 event in detail. We find that the highest predicted demand in the February 2021 event was well above the highest load observed in winter in that period. However, our estimate is in the range of observed extreme summer loads (Supplementary Note 1 and Supplementary Fig. 9).
Besides leading to high electricity demand, the low temperatures also caused substantial outages of generation capacities. As a result, loss of load occurred in 106 hours. Based on the predicted demand and the observed load, we estimate that in total, 1.45 TWh of load were affected by blackouts. Busby et al. 5 estimate the social cost of the power outages at US$130 billion. Therefore, the deficit cost of around US$87,000 MWh -1 is one magnitude higher than the value of loss of load used by the Texan market regulator ERCOT, that is, US$9,000 MWh -1 , in 2021.
Outages of gas generation capacities increased rapidly when the average temperature weighted by installed capacities at gas power plant locations dropped below -8.8 °C, which is a record low compared to the period 2004-2020 (Supplementary Note 1 and Supplementary Fig. 10). The outages were related to the freezing of power plants and of gas supply infrastructure, including production equipment at gas fields. Power plants outages increased rapidly when average temperature weighted by gas production at gas field locations dropped below -10.9 °C, a record low compared to the period 2004-2020 ( Supplementary Fig. 10). Therefore, gas supply infrastructure played an important role in causing the outages, as confirmed by ERCOT's classification of around 8 GW of gas power outages being related to limited fuel supply 9 . Coal generation capacity came offline at average temperatures weighted by coal plant locations of below -10.2 °C. This temperature is at the very lower end of observed temperatures in the period 2004-2021. For both coal and gas, recovery time was substantial. Even when temperatures increased to over 0 °C, 11.3 GW of thermal power plants-that is, 18% of total available thermal capacity-stayed offline for another 16 hours 10 .
At the time of the failure, temperatures at wind parks in Southern Texas were at the very lower end of the temperature range observed in the period 2004-2020. However, the average wind park temperature in Northern Texas when wind power plants started to fail was just below 0 °C and well within the range of previously observed low temperatures. On 13 February, when gas outages summed up to only 5 GW, ERCOT already reported 13 GW of wind capacity outages (Fig. 1). This represented a loss of 3.3 GW of wind power production on average at the prevailing wind conditions. However, later on, temperatures reached record lows at wind power plant sites in Northern Texas, too.
Our simulations of loss-of-load events using climate data from 71 years shows that the 2021 event was a record one. In total, we estimate that eight other severe power deficit events would have occurred in the current system if it had existed from 1950 to 2021, assuming the climate conditions of 1950-2021 (Fig. 2). The second largest power deficit event at 1.26 TWh is predicted when using climate data from 1983, assuming installed generation capacities as in February 2021. Furthermore, we observe 17 minor events. However, as the sum of the deficits of all 17 minor events is less than 1% of the sum of the deficits of the nine largest events, we exclude them from further analysis.
In our model simulations, the loss-of-load event has a duration of 106 hours and causes an aggregated deficit of 1.49 TWh, at a peak capacity deficit of 31.3 GW. There are several events with similar peak capacity deficits identified in the 1950-2021 period, and also events with a comparably long duration, but none with a comparably high amount of loss of load. In the largest events before 2021 (1962 and 1983), 250 GWh less lost load results from our simulation (Fig. 2). The year 1989 was the last time a similar freeze event occurred.
The 2021 record-high loss of load was not caused by the freeze magnitude alone but by a combination of a long, relatively cold freeze event and an inopportune timing of the freeze peak. According to Fig. 1, the system failure occurred early and was prolonged by a long freeze period afterward. This is in contrast to other years when   temperatures recovered more quickly after temperature minima had been reached (for example, in 1951 and 1963). This finding is supported by the extreme value statistics of the freeze spells shown in Fig. 3. The 2021 event was the second longest freeze event in seven decades. It has a return period of 37 years. Other events, however, were colder (1951,1989) or had higher frost sums (1951,1983). In terms of load, the highest predicted winter load in 2021 was slightly lower than the highest predicted winter load in the complete 71 year time series ( Supplementary Fig. 9), as there were lower temperatures in earlier years during the 1989 event.
Freeze events may have decreased due to global warming. However, it has been shown 11 that extremely cold events in the northern hemisphere have increased over the past 40 years. Our analysis for Texas does not indicate any significant trend in deficit events (Fig. 4) or freeze events (Supplementary Table 2). Still, average temperatures in Texas significantly increased due to climate change since 1951 (Supplementary Note 4 and Supplementary  Fig. 13). This result is confirmed by others; however, the increase in mean temperature is not genuinely transferable to extreme temperatures 12 . A stratified analysis of annual freeze events (minimum annual temperature) below temperature thresholds from 0 to -10 °C reveals that there is indeed no significant observed change of severe freeze events below -2 °C (Supplementary Note 5). Only very mild freeze events showed a significant attenuation (2.6 °C over the past seven decades), but such events are irrelevant to freeze-related failures of the power system comparable to the 2021 event.

Comparing revenues to cost of winterization
In the following, we assess how much revenue could have been earned by power generators by winterizing their capacities under  Average temperature weighted by population and predicted capacity deficit of severe freeze events within the 1950-2021 period. Labels in the graph refer to temperature minima and deficit maxima, which are circled: temperature (1a) and capacity deficit (1b) of the 1951 event; temperature (2a) and capacity deficit (2b) of the 1961 event; temperature (3a) and capacity deficit (3b) of the 1963 event; temperature (4a) and capacity deficit (4b) of the 1979 event; temperature (5a) and capacity deficit (5b) of the 1982 event; temperature (6a) and capacity deficit (6b) of the 1983 event; temperature (7a) and capacity deficit (7b) of the 1985 event; temperature (8a) and capacity deficit (8b) of the 1989 event; and temperature (9a) and capacity deficit (9b) of the 2021 event. Jan, January; Dec, December.
current regulation, assuming perfect competition and using the 71 past meteorological years. For that purpose, we use a chain of statistical and simulation models to derive loss-of-load events and revenues from winterization ( Supplementary Fig. 1). The revenues result from the scarcity price mechanism implemented by ERCOT: it increases prices automatically if spare capacity falls below a certain threshold 8 . The upper price limit is set to US$9,000 MWh -1 if spare capacity is below 2 GW ( Supplementary Fig. 8). If power operators winterize, they will be able to generate during times of high scarcity, generating additional revenue. Revenue from winterization is high but shows strong variability. For the first winterized gigawatt of gas power capacity, the expected revenue over a 30 year period is US$1.06 billion per gigawatt, but drops to US$0.52 billion per gigawatt at 14 GW of winterization (Fig. 5). Revenue for winterization of a coal power plant is slightly lower per gigawatt, and revenue for winterization of a wind power plant is substantially lower. For all technologies, the spread of revenues is high: the revenue at the 68% confidence interval is reduced or increased by half of the expected revenue. In 1.2% of all cases, there is no deficit event in a 30 year period, implying zero revenue from winterization. Substantial winterization measures can be implemented under our estimates of expected revenue, assuming that variable operating cost is low for power generators. This is true for nuclear, wind, photovoltaic and coal power plants. However, gas power plants may face high spot market prices for gas during cold spells, as observed during the 2021 event. We assume here that gas plant operators therefore have physically or financially hedged against high gas prices. We estimate that the winterization of gas wells-or building 250 GWh of pipe gas storage at gas power plant locations-in combination with winterization of gas power plants will cost about US$450 million per gigawatt (Supplementary Note 6). This cost is below the revenue up to the 15th gigawatt of winterized capacity. Winterization of coal and wind power plants is substantially cheaper, as no or very limited fuel supply infrastructure must be winterized. Winterization cost assumed at 10% of initial plant investments of coal power plants is far below revenues up to the full winterization of all failed coal capacity. In fact, one could assume that winterization costs 30% of initial plant investments, and even under this assumption winterization cost would be lower than the revenue for the completely winterized capacity. For wind turbines, our estimates of revenue are half those of coal, but are still substantially higher than the cost of winterization, which is reported to be 5% of investment cost 13 .
The assumed discount rate has a strong impact on results. When increasing the rate from 5% to 10%, the revenue for winterizing the first gigawatt of gas power plants is reduced from US$1.06 billion to US$0.65 billion. While winterization of coal and wind power still fully pays off under these assumptions, the profitable winterization of gas infrastructure and gas power is reduced from 15 GW to 7 GW.
Before 2021, these estimates might have been lower, as since 1989 no climatic event of a similar magnitude to that of 2021 had been observed. When dropping 2021 from our set of events, our estimates of expected revenue fall by 17.6% on average, causing only 13 GW of gas power to be profitably winterized. Still, when taking into account expected revenue, a substantial amount of capacity should have been winterized. Further assessment of the uncertainties, such as onset and recovery temperatures ( Supplementary Fig. 11), in our modelling approach can be found in Supplementary Note 2.

Potential reasons for the lack of winterization
We have shown that the Texas loss-of-load event in February 2021 was among the top three extreme events when simulating the power generation system under climate conditions of the last 71 years. Nevertheless, winterization of power generation infrastructure was already profitable before the event took place, taking into account the past climate, potential climate-change impacts and current regulatory conditions. So why did power generators not winterize? We identify several possible reasons for this gap between potential revenue and observed winterization efforts. First, while expected revenue is above winterization cost, its variance is high, in particular for gas power plants: for this technology, the winterization of the first gigawatt in 16% of all cases and the winterization of the tenth gigawatt in 36.2% of all cases results in negative profits. Additionally, owners of gas power plants have to secure their gas supply and may be exposed to high gas prices during cold spells. In contrast, the risk of not investing in winterization is low 14 . Even if a catastrophic failure occurs, the associated social cost is not born by the power generators. Potential costs include minor fines and damages to power plant equipment. A risk-averse investor may therefore decide against winterization. For owners of coal and gas power plants, another fact may reduce expected revenue: we assumed a 30 year lifetime for all power plants, but gas and coal generation infrastructure is partly old, and winterization may not pay off, if the plants are retired soon. For owners of wind power plants, however, we see the highest incentives and the lowest risk: the fleet is comparably young, winterization cost is low and revenues are substantially higher. Even with relatively high risk aversion, winterization seems to be a rational choice for wind power plant investors, especially if new wind parks are built.
Second, even though the historical temperature time series clearly indicates that the 2021 event could have been expected, the associated outages of power plants may have been underestimated by power generators. This is supported by King et al. 15 , who show that for a high number of plants, rated temperatures of failing power plants were in some instances substantially lower than prevailing temperatures during outages in February 2021. As no major outage such as the one in February 2021 had been observed before, plant owners may have assumed that current power plant standards were reliable enough.
Third, high discount rates also imply that winterization becomes substantially less attractive. At a 10% discount rate, our estimates of revenues drop by about 38.8%. Therefore, alternative investments with a higher return on investment and potentially lower risk may be preferred if limited capital is available, in particular for the more costly winterization of gas power generation.
Fourth, our calculation holds only under perfect competition. Owners of large generation assets have less incentive to additionally winterize ( Supplementary Fig. 12), as this would partly reduce scarcity prices for the already winterized part of their fleet. However, there is little concentration in the market in Texas, and we therefore see strategic behaviour as a potentially less important reason (Supplementary Note 3).

Conclusions
The total cost of extreme freeze events to society is at least one order of magnitude higher than the winterization cost. Winterization is therefore highly welfare enhancing. We have shown that current regulation provided a sufficient incentive for large-scale winterization for risk neutral investors, but risk aversion, lack of knowledge about potential outages or higher-yielding alternative investments may have impeded generators from making more effort to winterize. A more stringent regulation of winterization therefore seems to be necessary. In June 2021, Texas Senate Bills 2 and 3 became effective. They require the winterization of power plant infrastructure. However, gas power plants do not have to fully winterize; instead, a committee will define which installations have to 16 . As the failure of the gas power infrastructure by far had the largest impact on deficits in our simulations, we emphasize here that this impact should be strongly considered when defining rules for enforcing winterization of power plants and associated infrastructure.
Winterization of power generation units, however, may also come with downsides during periods of warm temperatures, as measures used to increase performance during cold weather, such as integrating gas turbines into insulated buildings, may make the cooling of power plants more complex. Apart from winterization, other mitigation measures, in particular strong demand response programs and an expansion of transmission capacities to neighbouring states, may therefore become important 6 . These mitigation measures may be substantially less costly, and they will be beneficial not only during cold spells. Both options can make the system more resilient against other variations in power generation, in particular taking into account the ongoing transition to a larger share of renewable energies in the power generation mix. During extreme freeze events, demand response may have to focus on industrial applications and less on households as electric heating in particular cannot be fully postponed during freeze events.
Of course, our results have to be considered in light of a continuously evolving power system. We assumed 30 years of lifetime for all installed capacities; however, some capacities, in particular coal power plants, may soon be retired, and therefore, their winterization may not be profitable. As the winterization of new capacity is cheaper and easier to implement than the winterization of existing capacity, winterization standards for installing new power plants and associated infrastructure should have a high priority. The ongoing transformation of the Texan power system can therefore be considered an opportunity to ensure robustness during future freeze events.

Estimation of temperature-induced electricity deficits.
To estimate the amount of deficits in the power system, we simulated the difference between the expected, temperature-dependent electricity demand and the available generation capacity. We used a regression model to simulate the demand from observations of past load in the years 2012-2020. Available generation capacity was obtained from the expected available capacity according to reports, reduced by outages related to temperatures.
According to ERCOT 17 , 67.5 GW of thermal capacity was available during the 2021 winter, but 4 GW of this may have been under maintenance. We therefore assumed a value of 63.5 GW of available thermal capacity before the freeze event.
Demand prediction. We predicted demand from winter temperatures using a regression model (equation (4) in the Supplementary Methods). To avoid the overestimation of loads at low temperatures with this model, a threshold temperature at -8 °C was introduced, under which the temperature dependency of the load was kept constant (Supplementary Fig. 2).
The model performed well for different temperature ranges in terms of average bias ( Supplementary Fig. 3), although at low temperatures we slightly underestimated the load. We therefore also ran the whole model chain with an alternative specification of the demand model where the temperature impact did not flatten off at -8 °C. The results are reported in our sensitivity analysis, indicating that the estimated deficits do not strongly change when using a different specification for the demand model (<10% difference). Testing the model out-of-sample for January 2021 delivered a high fit with a coefficient of determination (R 2 ) of 0.92 and a root mean square error of 1.31 GW. A cross-validation for other years (Supplementary Table 1) indicated a good fit.
Temperature-dependent generator outages. We estimated large-scale infrastructure failure aggregated by plant category. We derived the temperature, at which the largest increase in outages occurred for each power plant category, from the 2021 outage data (Supplementary Figs. 4-7). Furthermore, we estimated a constant outage level in terms of tripping capacity. Finally, we also defined a constant recovery rate, which describes how outages decrease after the recovery temperature is reached. This approach will omit smaller outages, but it accounts better for the inter-dependency of failures. Furthermore, the uncertainty in the data did not allow us to derive outage curves for individual plants. A detailed outline of how we derived the outage parameters from the 2021 data in Texas is given in the Supplementary Methods. The resulting outage models were applied to the 71 years of climate data to obtain the capacity availability during this period.
Estimating revenues from winterization. We determined the revenues from the scarcity price mechanism implemented by ERCOT 8 . This mechanism comes into effect whenever spare capacity falls below 8 GW. In this case, power market prices are regulated and set to fixed (high) values depending on the spare capacity. The lower the spare capacity, the higher the price, reaching US$9,000 MWh -1 when spare capacity falls under 2 GW (Supplementary Fig. 8). The difference between available capacity and demand was used to determine the current scarcity price and thus the revenue from winterization for each additional gigawatt of installed capacity for the 71 available weather years. The total revenue for a 30 year investment period with 5% discount rate was calculated. We did this for 10,000 scenarios, drawing randomly 30 years from the available 71 years to simulate different realizations of climate.
We emphasize here that this calculation only holds under a perfectly competitive market. For generators with large capacities that are already partly winterized, additional winterization may yield negative revenues, as market prices for existing winterized assets are reduced by the additional winterization. We discuss this in more detail in Supplementary Note 3.
Frequency analysis of freeze and power-deficit events. The frequency analysis of extreme events follows well-established methods of hydrological drought analysis 18 , where deficit events are defined as periods when the variable of interest is below a certain threshold. Here, we use two different threshold concepts. First, we analyse temperature and define a constant threshold of 0 °C to define deficit events, in analogy to drought events in drought statistics. Second, deficit events in the power system resulting from low temperatures are defined as periods when the capacity deficit is >0 GW (equation (1) in the Supplementary Methods). In each case, the result is a derived deficit time series, which is further investigated using Yevjevich's theory of runs 19 . During a freeze period, minor thaw episodes or other disturbances may split an event into several smaller events. As a remedy, pooling procedures have been recommended 20 . In this study, an inter-event time criterion of one day is used to define the deficit event series. In the case where multiple events occur in a year, the event with the absolutely largest accumulated deficit is used for further analysis. The derived series are characterized by three deficit characteristics: duration (measured in hours), intensity (minimum temperature and maximum power deficit for temperature and load deficit time series, respectively) and severity (aggregated frost sum and power deficit over the event for temperature and load deficit time series, respectively), each of which constitutes an annual extreme value series. These are further analysed using extreme value statistics to determine the return period of each freeze and capacity-deficit characteristic according to natural hazard management standards. Minor capacity-deficit events (with a duration of <6 hours) are excluded as these are not extreme events and would distort the extreme value modelling. Our analysis was conducted using the R-software package lfstat (ref. 21 ), which provides a collection of state-of-the-art methods that are fully described in the World Meteorological Organization's manual on low flow estimation and prediction 22 .
Data. The temperature at 2 m above ground is taken from the ERA5 reanalysis 23 . We calculate the average temperature over Texas weighted by population density 24 to derive a temperature index for modelling electricity demand. For estimating outages in the power system due to low temperatures, we derive the average temperature weighted by the capacities of wind 25 , coal and natural gas power plants 26 , which were the power generation technologies most affected by failure during the extreme temperature event of February 2021. For wind power plants, we split Texas into northern and southern regions (along the latitude of 30°), since temperatures at wind parks in the northern and southern regions of Texas differ substantially. Since the failure of the power system is also related to infrastructure at gas fields supplying these power plants 27 , we determine an average gas-field temperature index, weighted by the distribution of natural gas production by county 28 to complement our analysis. Load data used for demand prediction were retrieved from ERCOT 29 for the period January 2004-February 2021. Since the focus of this study is on freeze events in winter, only winter load data (December-February) is used. Outage data is provided in the period since 10 February 2021 by ERCOT 10 and is aggregated by power generation technology for the analysis.

Data availability
Aggregated climate data from ERA5 as well as results from the analysis, such as estimated load and threshold time series resulting from available capacity reduced by estimated outages and marginal winterization cost under different scenarios, are provided openly to the community on Zenodo: https://doi.org/10.5281/ zenodo.5902745. Data from public institutions, in particular ERCOT, the Energy Information Administration and the Texas Railroad Commission, are not available under an open license. However, within the description of the repository, links to data sources and the whole code, including download scripts, are provided so that our analysis, and in particular all figures, can be fully reproduced. Source data are provided with this paper.

Code availability
Code is published in a Github repository. The repository can be found at https:// github.com/inwe-boku/texas-power-outages.