I was at the end of my PhD, wrapping up my framework paper, when the power outages in Texas occurred. The past years I had been working on wind power simulation and validation, among other regions also in Texas. As large wind power generation capacities are installed in Texas and those were blamed for the rolling black-outs by some, this event was an excellent opportunity to apply the simulated time series in an analysis of this very recent event and make my thesis more practically relevant for a wider audience. I therefore started to download the relevant climate data the week after the event and asked my colleagues Peter Regner and Johannes Schmidt if they wanted to contribute.
Our working group is focused on economic power system modelling with a strong focus on understanding climate impacts. From joint teaching activities at the university, we were familiar with the statistician Gregor Laaha, who works on hydrological extreme events. He brought along Tobias Gauster. Both contributed with crucial analysis of return periods of extreme cold events and power system failures and the practical implementation of the statistical models in R. Together, we pondered on how to combine the methodology used for the analysis of hydrological drought events with capacity deficit events in power systems. Drought analysis uses a function which classifies run-off events as droughts if they are below a certain threshold. In contrast, we used the available power generation capacity as our threshold, and power demand as the equivalent of the run-off: whenever power demand was higher than power generation capacity, we classified this period as an extreme event.
The data problem
For the analysis of the Texas event we first of all needed appropriate data on grid load, and on outages in power plants. Obtaining the data was sometimes hard. Due to the recency of the event, many data sets were available only through social media platforms such as Twitter. Some data sets were only available in a non-machine-readable format like tables in pdfs. (A general recommendation at this point: please adhere to the FAIR Principles!). A further obstacle was the barrier imposed by ERCOT: access to ERCOT’s website is geo-restricted to internet users in the US. We had to take some detours to finally be able to download the relevant data sets.
Furthermore, the data was limited in its time span. Past load was available since 2004, but population weighted temperature never was as low during that period as during February 2021. This posed a critical problem to our demand model. In fact, we did not know how demand would react to very low temperatures and our model therefore seemed to overestimate the actual deficit. Luckily, when revising our study, we got access to ERCOT data on their load forecasts during the February 2021 event. This was crucial to improve our own demand model. The data set allowed us to adapt our model to very cold temperatures, implying a lower deficit than previously estimated.
Furthermore, shortly before submitting the article, ERCOT released a new, more refined data set on power plant outages, which made us redo our analysis. But finally, two months after the event, we were able to submit the first version of our manuscript to Nature Energy. After a thorough revision during summer following excellent reviews, we submitted our revised version by end of October. The paper was accepted (with minor revisions) end of 2021.
The core finding: Winterization is profitable, but…
With the help of our model, we estimated that eight similar events as the one in 2021 would have happened within seven decades, when using past temperatures, but the current configuration of the Texan power system. Taking into account scarcity prices in Texas, we were able to show that on average winterization of all assessed technologies is profitable. This also holds, for less capacities, if doubling the interest rate (from 5% to 10%). We also considered that the choice of the parameters in our outage model might impact results. However, a thorough sensitivity analysis on those showed that the number of events might change slightly, but profits largely still remain in a positive range on average. Furthermore, even assuming that climate change considerably changes the likeliness of extreme cold-events (which it does not, according to our analysis and other work), the profitability of winterization is reduced, but still significant.
So we wondered: Why were Texan power plants not winterized? Why were the recommendations following previous cold weather events impacting the power system ignored? We found one explanation in our results: The variability of expected revenues is very high (see Figure 1). Thus, there are non-negligible probabilities that an investment into winterization results in negative profits. For the winterization of the first GW of gas power plant capacity, the probability of negative profits is about 16%, and is increasing the more gas capacity is winterized. But this is not the only explanation.
Figure 1: Marginal costs and revenues of winterization for different power generation technologies: (a) wind power plants in the north and (b) the south (c) gas power plants, and (d) coal power plants. Costs of winterization include power plant (a-d) and gas field winterization (c). The confidence intervals are the result of 10,000 draws of 30 year investment periods from 71 years of climate data. Revenues are determined from the difference in profits due to winterization using the current scarcity price according to ERCOT, which depends on the spare capacity available.
Looking for explanations for the lack of winterization
Power generators do not have to bear the cost of rolling blackouts. So they face an up-front investment cost for winterization, which may not pay-off, but no loss if not winterizing. For loss averse investors, this may imply under-investment into winterization.
Another explanation might be that discount rates are way higher than we assumed and therefore investment into winterization is less profitable. Similarly, investments don’t pay off when power plants are comparably old and the remaining lifetime is short: the likeliness that no extreme event that allows to earn back investment cost occurs during the remaining life-time is high in such a case.
It is also possible that power suppliers simply underestimated the effect of the cold temperatures and did not anticipate such catastrophic failures. This is confirmed by reports which show that rated temperatures for power plants were significantly lower than observed temperatures when outages occurred. However, earlier events e.g. in 2011 already had shown that freezing weather can have a significant impact on the Texan power system.
What lesson can we learn from this?
To sum up, there are several possible reasons that might have made investment into winterization measures unattractive in the Texan power system - but the risk associated with winterization events may be the most important one. The current incentive structure leads to market failures with severe consequences for Texan consumers. These market failures need to be corrected either through direct regulatory intervention, such as mandatory winterization of infrastructure, or by creating appropriate risk markets. In particular, gas power plants and the gas infrastructure need to be targeted, as they currently provide the largest source of dispatchable power capacity in the system. While this event was a highly interesting opportunity to better understand the performance of energy-only markets under extreme events and expand our research, we sincerely hope that such events can be avoided in the future.