New Killer Star: Implications of the COVID-19 Pandemic for Climate Change and the Fossil Fuel Industry

I've written before about climate change. That page presented an analysis of numerical data pertaining to the topical topic of climate change/global warming. The results are presented with Tableau. As a prelude to what I've done more recently, one of those dashboards is included here.

The motivation for what I did then was the tension between scientific consensus on the allocation of responsibility for climate change and divided public opinion on the matter. The visualisations, particularly the interactive modelling dashboard, were an attempt to show how clear the answer could be made. As far as I knew, this had not been done as clearly before.

The COVID-19 pandemic made me curious about how that would affect the likely climate outcomes. Climate outcomes are a strong function of the resources needed to change course if necessary, and resources available, with the latter potentially severely impacted by the pandemic and its global political effects. So the focus of what I do here is different - leaving public opinion to one side, what is actually likely to happen? And how much would it take to change course, if necessary?

The COVID-19 pandemic has resulted in a lot of global uncertainty, both because of the direct economic disruption and the associated global political fallout. When we look back on the twentieth century, its defining events and circumstances are the world wars, great depression and cold war. We can't yet know what events and circumstances the historians of the future will list when they look back on the twenty-first century, but two strong candidates are the COVID-19 pandemic and climate change, as true global disruptions. Part of the motivation here is how, if at all, these two events will interact with each other. We do not simulate the climate system as rigorously as government agencies can, but we do this more quickly, and with flexibility to explore the particular implications that interest us most.

Getting down to specifics, there has been speculation about the potential silver lining of the COVID-19 cloud. With economic activity slowing significantly, commensurate slowdowns in heavy industry and transport - sources of greenhouse gases (GHGs) - are expected to lead to the most dramatic declines in GHG emissions in decades, perhaps ever. The pandemic is killing people, but can we at least claim a consolation prize: a solution to our climate change problem? And with the prices of fossil fuels like oil and coal falling dramatically, what will happen to the fossil fuel industry? It's certainly been wounded, but will that be enough for renewable sources of energy to take significant market share in the energy sector?

The markets have never priced fossil fuels accurately, according to the actual burden that they impose on the environment. The fossil fuel industry has enjoyed the benefit of untaxed externalities. So up until this point, governments have had to intervene in the energy sector to give renewables a chance - with regulation, taxes and emissions trading schemes. But do renewables still need a leg up? Put another way, due to the disruption caused by the pandemic, need governments continue to level the playing field for renewables? This question goes back to the observation above, about the increasingly-uncertain international political environment - with governments allocating more resources to national security, they might prefer to address those priorities.

Governments would be interested in these considerations, partly because they are responsible for approval of energy sector assets like power plants. They are also collectively responsible for action on climate change - and the pandemic's disruption might change this cost. Anyone operating in the energy sector would also be interested.

On this page we consider the impact of the COVID-19 pandemic on climate change, whether renewable energy price reductions along with market forces will be enough to avert climate catastrophe, and if not, what it might take to change course. The page is divided into 6 sections.

In section 1 we a dashboard from the earlier analysis of climate data - the one about principal component analysis, which remains of interest in the present work. Following this, in section 2 we train a temperature anomaly estimator and explain how it is used as part of a simulation to predict warming over the century. In section 3 we describe the simulation that forms the context in which the estimator operates, and that includes effects of the pandemic. In section 4 we describe effects built into the simulation, most of which relate to the impact of changes in the fossil fuel industry, which have probably been accelerated by the pandemic. In section 5 we perform Monte Carlo simulation to find probabilities of avoiding particular warming thresholds. In section 6 we consider the effect of solar energy subsidy. In section 7, code is provided and described. In section 8, credit is given where due.

A coarse-grained plan:

What we did back then was look at the warming, visualise anomalies that might have led to it and fit models that might explain it. We found that a decent fit could only be obtained if anthropogenic forcings are considered. On this dashboard we examine just how important each forcing is likely to be in the cause of warming.

With a Principal Component Analysis (PCA), we reduce our 8-dimensional parameter space (over 1984-2012) to just two principal component dimensions, while retaining 77% of the dataset's variance. The percentages in the principal component axis titles are how much of the dataset's variance resides in each principal component. The original data are plotted in this principal component space year-by-year. The original parameters are plotted as radial vectors; a smaller angle between two parameter vectors indicates that they tend to vary more together. These vectors are plotted on the component axes.

In the principal component space, temperatures and the anthropogenic forcings all increase to the right. This shows that they tend to rise and fall together. Vectors for the natural forcings are essentially perpendicular to the others, so they have little to do with either human activity or with the overall trend in temperature. These inferences are confirmed by the projections shown, of forcings onto temperature and human population. The most extreme data point, 1992, followed the eruption of Mount Pinatubo in 1991. That did signifiantly affect temperature, but the natural forcings can only help to explain year-to-year temperature variation, and some trends over periods of a few years - not the overall trend.

This is highlighted most clearly by the plot in the bottom left. Here we show the contributions of each forcing to warming, according to ANModel. Earlier in the project we spelt out the terms of this model in a formula. The contributions of specific forcings to warming are the corresponding terms in that formula. The natural forcings all oscillate, in some years causing cooling relative to the base period. The sense in which stratospheric aerosols ever lead to warming is indirect: when there are no significant volcanic eruptions, there is effective warming, relative to the base period.

If that project has a conclusion, it might be the column chart comparing the anthropogenic and natural contributions to warming.

Here we compare the contributions of anthropogenic and natural forcings to warming, which vary over time. These are averaged contributions over the period of time set by the sliders on the right. To see the average contributions of anthropogenic and natural forcings for a particular timespan, type the start and end years, or use the sliders.

There is one year in which, according to this, the natural contribution to warming was greater: 1987. But no others. If we look at 2010 by adjusting the start and end sliders to that year, the combined warming from anthropogenic greenhouse gases is 0.5 degrees. This is consistent with (at the bottom end of) the IPCC's assessed likely range for warming over the 1951-2010 period attributable to well-mixed greenhouse gases (Figure 1.9 in the IPCC's Climate Change 2014 Synthesis Report).

2. Temperature Anomaly Estimator

    Heart's filthy lesson
    Falls upon deaf ears
    Falls upon dead ears
    - David Bowie, "The Hearts Filthy Lesson" (1. Outside)

What the PCA tells us is that to predict the temperature anomaly, the most important features from the models tested are the GHGs - essentially, AModel. The natural features have some predictive potential, but the additional accuracy they could provide would be marginal, especially if we are interested in large timescales, over which there will be more significant uncertainties. Moreover, while we have some idea how GHG concentrations may vary over this century - via their covariance with human population - most of the natural forcings exhibit periodicity that has no perceptible trend. Put another way, GHG concentrations are enough to help us predict the trend of warming, which the natural forcings do not. This trend, rather than the precise temperature anomaly in any specific year, is what we are trying to predict. It is the anthropogenic forcings that can help us to do this, so these are the features we proceed with.

So our features to predict the temperature anomaly in any given year are that year's carbon dioxide concentration, methane concentration and nitrous oxide concentration. These are well-mixed gases, with the mixing as a process that takes up to three years. For this reason, we technically use three additional features (for a total of six), one for the concentration of each GHG in the year before. Since the GHG concentrations are increasing with time, predictions for future temperature anomaly rely on values outside the training data space. So KNN and tree-based algorithms fail; some algorithms that work are gridsearched, leading to these results:

Three of the four estimators are in excellent agreement, and fit the trend nicely. Boosting overfits, and all other estimators have R² = 0.81. They would probably return very similar predictions, but linear ridge regression has a marginally better CV score than the others. In the previous project we found that a linear model went well, so this isn't surprising. A more basic reason that this shouldn't surprise is that we are looking at (relatively) small perturbations from an equilibrium state of the climate system. We certainly don't expect the temperature anomaly to generally be a linear function of the GHG concentrations, but for small perturbations from equilibrium, it's plausible.

3. Simulation Environment and Overview

    (Time will crawl) And our heads bowed down
    (Time will crawl) And our eyes fell out
    (Time will crawl) And the streets run red
    (Time will crawl) Till the 21st century lose
    - David Bowie, "Time Will Crawl" (Never Let Me Down)

To predict temperature anomaly over this century with our estimator, we need to first estimate GHG concentrations over that time. These can be crudely predicted with a series of linear fits between population, GDP, emissions and GHGs. GHG concentrations increase with emissions, which in turn result from human activity. A proxy for human activity is GDP, and this is related to population. So we perform a series of fits to historical data: population-GDP, GDP-emissions and emissions-concentrations. We take the coefficients from these fits and use them to predict, based on predicted population, how GHG concentrations will vary over this century.

We can't integrate these fits into the temperature estimator because:

• emissions data are not as complete as concentrations: we'd have to train on even fewer than 33 years of data
• we're entering territory that is uncharted in our training data, with elements that need to be treated separately from the estimator - there were no pandemics in the last thirty years

We obtain population predictions from the UN, GDP numbers from the World Bank and GDP post-pandemic predictions from IMF. GHG emissions data are also from the World Bank, and GHG concentrations for the last three decades from various sources (often NASA). Temperature anomaly data are from HadCRUT. Our sources are listed here.

The GDP prediction is not a simple fit. We first fit GDP per capita to time, and then multiply this by population to predict total GDP. This is the base prediction - GDP, as well as other time-varying parameters, is modified by other effects throughout the simulation, as we describe below.

From a data standpoint, the unprecedented nature of the pandemic is part of what makes a simulation necessary. The effects of the pandemic are many, and we don't include all of them. The ones we do include are:

The various parts of the simulation are described below.

4. Simulation Effects and Fossil Fuel Industry

    Ya wouldn't believe what I've been through
    You've been so long
    Well it's been so long
    I've been putting out fire with gasoline
    Putting out fire with gasoline
    - David Bowie, "Cat People" (Let's Dance)

Our simulation discretises time, such that the various quantities are calculated for each year of the simulation. The most important of these quantities are shown in the diagram above. Time is run forward from 2020 until 2100. The temperature anomaly estimator is used within the simulation at each time step, to predict just one temperature anomaly.

Effect: Temperature Anomaly Hurts GDP: We take the predicted temperature anomaly from each time step and reduce GDP for the next time step. We use Kalkuhl & Wenz (2.0-2.9% per degree of warming), but the estimated impact varies considerably. It's expected that the impact of warming does not increase linearly. This has not been incorporated into our simulation. A figure from NGFS Climate Scenarios showing the variation of estimated impacts is:

Effect: Lifetime of Greenhouse Gases (GHGs): The persistence of different GHGs in the atmosphere varies considerably. Carbon dioxide (CO₂) gets "top billing" not because it is the most potent, but because it tends to linger in the atmosphere. Concentration half-lives are reported as lifetimes by the IPCC.

Figure 8.29 from the 2013 IPCC report:

12.4 years for CH₄, 121.0 years for N₂O. CO₂ is complicated but we have used 100 years.

Pandemic and GDP: The pandemic reduces GDP, and we can use the GDP-carbon emission correlation to adjust the predicted carbon dioxide concentration. If we assume that GDP is 8% lower during the pandemic than it would otherwise be, it makes no significant impact on warming, even for a 10-year pandemic. The temperature anomaly in 2100 is different, but only by about a hundredth of a degree. This is because carbon dioxide persists in the atmosphere for a century or more, and the effect of the reduced emission is "smeared out" over that time. Some hard numbers can make this easier to understand: if the temerature anomaly is determined by emissions over the preceding hundred years, a 10% reduction for 10% of that hundred years would only reduce the change by about 10% of 10%, or 1%.

Pandemic vs Fossil Fuel Industry: The pandemic has delivered an immediate hit to GDP, causing a deep recession. The fossil fuel industry has been expanding until now, but it requires investment: capital expenditure to develop fuel reserves. So while almost all industries are vulnerable during a recession, the fossil fuel industry is especially vulnerable. The COVID recession means less money for investment, and what money remains is likely to go where there is less risk. Reduced demand has led to falling fossil fuel prices (oil prices have gone negative, coal price has fallen from $80 in 2019 to $50 in July 2020, in US dollars per tonne). The fossil fuel industry is really not the place to be for the risk-averse.

The industry is likely to get real about the costs: As much anything else, the recession has forced a critical re-evaluation of the fossil fuel industry's future. The best evidence that it hasn't been doing this is that coal power stations were still being built just before the pandemic - even though they've been less competitive than solar (and wind) farms for at least 5 years. Getting real about this will, among other things, make the industry more predictable, which helps us.

Renewable Alternative? But challengers to conventional coal power face an uphill battle. Coal is proven technology, the plants are already built and we know how to scale the technology up for mass-reliability. Any challenger needs to go way beyond technological proof-of-concept. It needs to compete with established infrastructure, not just physical, but economic and political as well.

The Cost of Being the Challenger: Fortunately, we can quantify the cost of being the challenger. A challenger that lacks infrastructure needs to pass this cost on, making it less competitive. So for the challenger, it's the Levelised Cost of Energy (LCOE) that counts, which is the cost that counts establishment - in large part, construction of the power plant or farm. In most markets, because coal is established, challengers compete against the Marginal Cost (MC) of coal power, not its LCOE. MC is much less than LCOE for any power generation method, so incumbency helps coal.

The Comparison That Counts: If we were comparing energy sources with a similar measure (LCOE), we would have stopped burning coal for electricity some time ago. But because it's the challenger, solar PV has its LCOE pitted against coal's MC. A severe handicap, but it is starting to become competitive on that basis.

Energy costs here are from Lazard:

Question: Will Solar LCOE Beat Coal MC? We can fit a decaying exponential to the solar LCOE:

• LCOE = 325e^{-0.395t + 34}

The constant term implies that it approaches $34, which it reaches by 2030. This is basically the typical marginal cost of coal electricity. So if this is what happens, in at least half the markets, we'd need to wait for coal power stations to reach end of life, and the usual lifetime of a coal power station is fifty years.

Early Closure of Coal Power Stations: For a significant fraction of markets, it looks like coal power stations could be closed early because running them would be more expensive than building new solar farms instead. We find this fraction by comparing the predicted solar LCOE with the range of coal marginal cost. Temperature anomaly is about half a degree lower by 2100, and rise is delayed by a decade or two.

Effect: End-of-Life Closure of Coal Power Stations: If we assume a new rationality, very few new coal power stations will be built. We also assume that all existing coal stations will be closed 50 years after they began service. We can (crudely) predict the curve of (maximum) coal power generation from historic output (using coal power from BP).

With the pandemic, coal power station projects in the developing world have been getting cancelled. We simulate this by bringing forward the peak coal power generation capacity from 2030 to 2020. Emissions are adjusted to match this decline to zero fifty years from now, in 2070. But this effect is slight, because construction of new coal power stations was slowing anyway.

Summary of Pandemic Effects on Temperature Anomaly: A 10-year pandemic makes a difference, but only a few hundredths of a degree in 2100. In 2100, the temperature anomaly is 2.2 degrees. The temperature rise that people worry about is relative to pre-industrial times, and this would reach about 2.5 degrees. Not quite "global hot house", but would trip some tipping points.

The reasons for the pandemic not having the effect on GHG concentrations in the manner we might expect are:

• the temperature anomaly is a function of emissions over about a century
• the changes in the energy sector were happening anyway - the pandemic only accelerated them

But so far, we have expectation values with no probabilities or uncertainties! We run the simulation a thousand times with parameters varying like this:

Parameter	Values	Source	Type of Uncertainty
Temperature Anomaly Estimator Coefficients	(Many)	(Procedural)	Statistical - Bootstrap
GDP Growth per year	1.0-2.8%	Leimbach et al. 2017	Statistical - 2σ Gaussian
Pandemic Hit to GDP (in each year)	4-11%	IMF 2020	Statistical - 2σ Gaussian
Warming Hit to GDP (per degree)	2.0-2.9%	Kalkuhl & Wenz 2018	Statistical - 2σ Gaussian
Atmospheric Carbon Dioxide Effective Half Life	100 ± 50 years	IPCC 2013, Figure 8.SM.4	Statistical - 1σ Gaussian

To estimate the uncertainty in the underlying climate model, we:

• take a thousand random subsets of the original data
• fit a linear estimator to each
• generate an ensemble of a thousand estimators to include in the whole Monte Carlo ensemble

The result when these estimators are compared with the past few decades looks like:

For all other parameters, we generate Gaussian distributions with the required central tendencies and spreads. The simulation is run once for each set of parameter values.

Probability of Avoiding a Temperature Rise of:

	1.5 Degrees	2 Degrees	3 Degrees
By 2050	0.863	0.967	0.992
By 2100	0.033	0.233	0.847

Applying our crude model of energy pricing and market share, in which the fraction of market share taken by solar is directly related to the amount by which its LCOE beats coal's MC on price, we find that coal would be replaced completely by solar if it was subsidised at a rate of US$15/MWh. We need to remember, though, that this won't stop the warming, both because the carbon dioxide already in the atmosphere would hang around, and there are other GHGs that we're not modelling action on. But as far as the effect of coal goes, we can alter the subsidy for solar and watch how the effect on carbon dioxide changes the fraction of simulated futures in which the various temperature thresholds are crossed.

A US$15/MWh subsidy might increase the probability of avoiding a 2-degree temperature rise by 2100 from about 20% to 30%. The world's total primary energy supply in 2017 was 162 494 TWh, and a thousand times as many MWh. With this as an upper limit, we can calculate that the cost of subsidising to ensure all of those MWh come from solar would be about 2.4 billion US dollars. If we did that right now, the cost for this year would only be 2.4 billion US dollars. The impact of a 2-degree temperature rise in 2100 would be at least 5% of then-GDP, perhaps 25 trillion US dollars.

Probabilities are the numbers we want if we're curious about how likely particular futures are. But arguably, a better way to quantify the return on investment is by examining the change in the expected values. The most likely temperature rise by 2100 falls to 2.3 degrees. Considering the impact on GDP expected by Kalkuhl & Wenz 2018 with a conservative estimate of world GDP by then (US$500T), there might be a saving of about 2.5 trillion dollars . . . per year!

So it's pretty clear that an investment of billions per year now will save trillions per year later. The final calculations to reach those figures are crude, but what uncertainties remain unaccounted for would need to be huge to change the obvious advice - that solar energy is worth subsidising.

7. Code

As I write this, I still haven't got around to posting the code here. But when I get a moment . . .

Name	Purpose	Requires	Language
load_climate.m	Load climate data: GHG concentrations, temperature anomaly. Create feature arrays.	The data mentioned.	Octave
prep_covid.m	Load population data, perform fits for population-GDP, GDP-emissions and emissions-concentrations. Early version of the prediction that I ended up doing differently.	The data mentioned.	Octave
coalcapacity.py	Predict coal generation capacity, on the assumption that no more coal poer stations will be built and te lifespan of a coal power plant is 50 years.		Python
exploration_covid.py	Mostly exploration, but also collation of GHG data for eventual inclusion in features, not that I do it this way anymore.	Depending on the options selected, may require: load_covid.py, prep_covid.m, exploration_covid.py, load_climate.m, learn_climate.py, show_climate.py, powercosts.py, coalcapacity.py	Python
learn_climate.py	Conventional machine learning with various algorithms.		Python
load_covid.py	Load emissions and GDP data.		Python
mkmcmodels.py	"Make Monte Carlo models:" train many estimators on random subsets of the original data, and store these estimators for inclusion in the Monte Carlo ensemble		Python
nks.py	Central script that runs many of the others. Loads all of the data so the workspace is ready to train estimators.		Python
powercosts.py	Makes all of the plots of LCOE and MC for coal and solar power.		Python
show_climate.py	Plots the fits achieved by the various estimators to data from the past three decades. Also plots the fits achieved by the Monte Carlo ensemble of estimators.	Estimators trained by learn_climate, or mkmcmodels.py.	Python
sim.py	Simulation. Also generates ensemble parameter sets.	A temperature anomaly estimator either in the workspace or saved from learn_climate.py. mkmcmodels.py, or estimator ensemble saved by mkmcmodels.py. Saved ensemble parameters, unless set to generate a fresh set. Coal power generation capacity from coalcapacity.py.	Python
simresults.py	Plots results from simulation, which are usually saved as pickle files	Pickle files from sim.py	Python

8. Acknowledgements
Data Sources

Data	Researchers and/or Organisations with External Link	Local Link
Human Population - Prehistoric	Netherlands Environmental Assessment Agency HYDE database, via Our World in Data	Here
Human Population - Recent	United Nations Department of Economic and Social Affairs, Population Division, via Our World in Data	Here
Temperature Anomaly by Location	Climatic Research Unit (University of East Anglia) in conjunction with the Hadley Centre (UK Met Office), via Our World in Data	Here
Temperature Anomaly - Prehistoric	Shakun et al. (2012), in Nature	Here
Temperature Anomaly - Recent	Climatic Research Unit (University of East Anglia), Hadley Centre (UK Met Office)	Here
Carbon Dioxide - Mauna Loa	IPCC	Here
Carbon Dioxide - Global Mean	Ed Dlugokencky and Pieter Tans, US NOAA/ESRL	Here
Carbon Dioxide - Prehistoric	J.-M. Barnola, D. Raynaud, C. Lorius, N.I. Barkov. US CDIAC	Here
Methane	Ed Dlugokencky, US NOAA/ESRL	Here
Nitrous Oxide	US AGAGE	Here
Solar Irradiance	Judith Lean, US NASA GISS	Here
Stratospheric Aerosols	Makiko Sato, Andrew Lacis, James Hansen, Larry Thomason. US NASA GISS	Here
Southern Oscillation Index	Australian Bureau of Meteorology	Here
World GDP, emissions	World Bank Open Data
Electricity prices in the US market	Levelized Cost of Energy and Levelized Cost of Storage 2019, Lazard
World Energy Statistics	BP

The earlier analysis I did can be found here. The data I used, along with code I wrote, can be found here. Numerical analysis (apart from the basic computations) was carried out with Python (plus numpy, pandas, scikit-learn, oct2py, matplotlib and plotly) and GNU Octave. The title of this work is also a David Bowie song, "New Killer Star" (Reality).

Illustrations on this page: Stills from the "New Killer Star" music video.