Detecting Bitcoin Miners By Their Carbon Emissions
Can we identify power plants used for cryptocurrency mining using only their carbon emissions? Yes we can, read on to find out how.
Researchers from the Columbia University Center for Environmental Economics and Policy (CEEP) recently published a paper titled “Bitcoin and carbon dioxide emissions: Evidence from daily production decisions”. Thanks to Maximilian Auffhammer for highlighting this in the Energy Institute at Haas blog.
The researchers picked a coal power plant in Pennsylvania that is exclusively used to power Bitcoin mining, and investigated whether carbon emissions from that plant were influenced by the market price of Bitcoin. Here’s their overall conclusion:
We find that carbon emissions respond swiftly to mining incentives, with price elasticities of 0.69–0.71 in the short-run and 0.33–0.40 in the longer run. A $1 increase in Bitcoin price leads to $3.11-$6.79 in external damages from carbon emissions alone, well exceeding cryptomining’s value added (using a $190 social cost of carbon, but ignoring increased local air pollution).
In other words, when the price of Bitcoin goes up, the owners of this coal plant crank it up to mine more Bitcoin, producing more carbon emissions as a result. They included this chart comparing the price of Bitcoin over time to the tons of carbon released from this plant.
As the CEEP paper notes, “data on US Bitcoin miners are […] sparse”. I wondered: could we identify more plants used for mining by comparing their emissions to the price of Bitcoin?
I used data for January 2018 through the present, similar to the time period analyzed in the CEEP paper. I downloaded monthly emissions data from the EPA’s Clean Air Markets Program Data, which contains physical measurements of actual CO2 emitted by power plants, and the monthly closing price of Bitcoin from Yahoo Finance.
We want to find power plants with an emissions time series that closely matches the Bitcoin price time series. We’ll use the Euclidean distance metric to quantify the similarity, after normalizing all the data to fit the same scale. If you don’t care about the math, skip ahead to the Results section.
Normalization
Since the emissions and pricing data are on very different scales, we’ll apply two transformations to normalize the data. First, subtract the mean:
mean = statistics.mean(data)
data = list(map(lambda x: x - mean, data))This removes irrelevant absolute differences between time series — if the pricing data is averaged around 30,000 and the emissions data is averaged around 5,000, they’ll now both be averaged around 0 and be more easily comparable. Next, we’ll apply amplitude scaling:
mean = statistics.mean(data)
stdev = statistics.stdev(data)
data = list(map(lambda x: (x - mean) / stdev, data))If a $1 increase in Bitcoin price corresponds with a 5x increase in emissions, we want to treat that the same as a 1-to-1 correspondence between price and emissions. The question we’re asking is “does emissions increase and decrease with price”, we aren’t concerned with the magnitude of that increase or decrease. Amplitude scaling masks this difference.
This is just the bare minimum of normalization. There are other techniques we could apply, such as detrending: if a mining facility is continually adding new computing power over time independent of the price of Bitcoin, we could identify that trend and remove it. We could also try smoothing the data to reduce irrelevant noise.
Results
We can now easily compute the Euclidean distance between two normalized time series:
numpy.linalg.norm(normalized_emissions - normalized_prices)I ran this comparison for every plant in the EPA dataset. Let’s look at the first few plants with the highest similarity to Bitcoin prices.
First on the list is #10143, a coal plant in Pennsylvania.
False alarm. This looks like a plant that was completely shut off and coincidentally started back up around the same time that Bitcoin spiked. We could easily filter out plants like this.
The next closest match is #335, a methane-powered plant in Huntington Beach, California.
Suggestive, but no clear evidence — this plant is owned by AES, a large utility and power generation company. However, there is Bitcoin mining in Huntington Beach, including Nate’s Food Co, which claims to be a failed pancake mix business that pivoted to Bitcoin mining (???).
“Nate’s Food Co. is engaged in Bitcoin Mining. The Company owns application-specific integrated circuit (ASIC) computers specifically designed for cryptocurrency mining. […] The Company is also focused on licensing its developed products consisting of a ready-to-use, pre-mixed pancake and waffle batter delivered in a pressurized can.”
The Nate’s corporate office is less than 10 miles from this plant.
Next up, #55749, a coal plant called Hardin Generating Station in Montana.
Jackpot! Bitcoin miners purchased the Hardin plant in 2020, which almost certainly would have closed for good without their intervention.
Next, #2527, the Greenidge Generation methane plant in New York.
Another hit! “Greenidge is the first vertically integrated power generator and Bitcoin miner of scale in the United States.” The Greenidge plant has been deeply controversial, and last year was found to be out of compliance with New York State environmental law.
Other confirmations will be harder to come by, as most miners are not transparent about where their power comes from and may not own their own plants. Many other plants displayed a high correlation to Bitcoin, but I was unable to find additional evidence. Looking up plant ownership isn’t enough to make a determination — even old-fashioned utilities have been known to enter the Bitcoin game. I’ll highlight one more example case that warrants further investigation.
Below are three methane gas plants in Southern California (#55510, #55513, and #55540). The first two, which exhibit a strong similarity to the Bitcoin pricing data, are owned by Calpeak Operating Services, LLC. The third is owned by a different company and shows no correlation to Bitcoin, despite being a similarly-sized plant located close to the other two.
The Calpeak plants are labeled as peaker plants, which are typically only called upon by grid operators during periods of high demand. Peaker plants are capable of turning on and off very quickly, but at the price of decreased efficiency — they’d make more money if they could run all the time. Could Calpeak be using the excess capacity of these plants to mine Bitcoin? Maybe! Allow me to speculate wildly: Calpeak is owned by Middle River Power, which is itself owned by Avenue Capital, a private equity firm. Maximizing the revenue of power plants by mining cryptocurrency would certainly be in character for PE firms, which are typically focused on short-term financial results, and Avenue Capital is owned by Marc Lasry, a public proponent of Bitcoin.
This rudimentary analysis demonstrates that power plants known to be associated with bitcoin mining can be easily identified by their CO2 emissions alone, and further analysis could potentially be used to identify previously unknown miners.
If you’re interested in this topic, I recommend reading the full paper that inspired this post. If you’re interested in learning more about carbon emissions and the electrical grid, check out Singularity Energy, and use our free API or open-source emissions project to dig into the data.
