Water, Wind, and Fire:

The Paris Agreement and the Cost of Energy Transition∗

Rodrigo Moita†, Daniel Monte‡, and Victor Orestes§

August 29, 2020

Abstract
The Paris Agreement is a global climate change agreement in which many countries took
on significant emission cuts. We develop a dynamic model of electricity markets with ther-
mal, hydro and intermittent power sources to simulate the costs of fulfilling the agreement.
We calibrate our model to data from Argentina, Brazil, and Uruguay, three countries with
very different capacity matrices and diverse natural characteristics. We compute the costs
of reaching a (close to) zero emissions and find them to be quite significant, but with large
differences among the countries. We also analyze the effect of the recent trend of falling prices
of investment in renewable capacity. The results show that this is not enough to move to a
clean energy matrix.

JEL codes : D41, D47, Q41, Q54

Keywords: Climate change, electricity markets, hydroelectric generation, intermittent sources

∗
We have greatly benefited from discussions with Luciano Castro, Bernardo Guimaraes, Richard Hochstetler,
Shaun McRae, Leonardo Rezende, and Marek Weretka. Financial support for this study was provided by Aneel’s
RD program PD-0678-0314-2014 sponsored by EDF Norte Fluminense, EDP and Energisa. This paper was partly
developed under the project "The Regulation of Public Utilities of the Future in Latin America and the Caribbean",
for the IADB and we are very grateful for their support and for the many comments and suggestions that we received.
†
USP. E-mail: rodmoita@usp.br
‡
Sao Paulo School of Economics - FGV. E-mail: daniel.monte@fgv.br
§
MIT. E-mail: orestes@mit.edu

1
1 Introduction
Climate change is one of the most serious issues affecting the world today. Reaching a multilateral
agreement that could ultimately lead to a sustainable growth path proved to be very difficult,
both technically and politically. The most recent of such agreements was finally reached in 2015.
The Paris Agreement is a universal global climate change agreement, adopted at the Paris climate
conference (COP21) in December 2015.1 The main goal of the agreement is to keep the increase
in the global average temperature to well below 2◦ C above pre-industrial levels and to limit it
to 1.5◦ C. The agreement is considered successful in the sense that many more countries took on
emission cuts in the Paris Agreement than in its predecessor, the Kyoto Protocol.2
However, despite this wider embracing and an optimism with the novel bargaining procedures
and enforcement mechanisms in this agreement Harstad [11], achieving its goal is a daunting
task. In order to fulfill its commitment under the Paris Agreement, the European Union aims to
be climate-neutral by 2050 - an economy with net-zero greenhouse gas emissions. This implies
reducing carbon emissions as soon as possible, which in turn requires a structural change in the
electricity production sector, which is the number one industry in terms of carbon emissions.3
The electricity industry has faced important structural changes in the last three decades. There
has been a wave of liberalization, privatization and competition in the energy generation sector.
New technologies are entering the energy system and intermittent-renewable sources such as wind
and solar now account for a significant share of the generation in many countries. The Paris
Agreement is accelerating these changes. The ambitious goal of zeroing emissions from electricity
generation until 2050 implies an ever increasing share of renewable sources, as well as a decreasing
share of thermal plants. Renewable sources are cleaner and have lower marginal cost, but they
lack the reliability of the thermal generators. The absence of rain, sunlight and wind may imply
a shortage of electricity supply. So, in this context, central questions are: how to achieve the
Paris Agreement goal without compromising supply reliability? What is the cost of achieving this
goal? Falling investment cost of renewable technologies are enough to significantly lower carbon
emissions, or do we need public policy?
It is important to develop a specific framework to study these questions, since electricity markets
function in a quite different way than other commonly studied markets.4 There are several features
specific to this industry: for the most part, electricity is non-storable and non-tradable (outside
the grid), supply and demand vary substantially across time and the total potential supply is
1
It is an agreement within the United Nations Framework Convention on Climate Change (UNFCCC). See
European Commission [4].
2
As of February 2020, all UNFCCC members have signed the agreement.
3
To exemplify this need for a significant structural change, consider the impact in global industrial greenhouse-
gas emissions caused by the Covid-19 crisis. The International Energy Agency expects that emissions will be at most
8% lower in 2020 than they were in 2019 (https://www.iea.org/reports/global-energy-review-2020). Since warming
depends on the cumulative emissions to date, this large drop in 2020 hardly makes any difference in achieving the
Paris Agreement’s desired goals.
4
See, for example, Joskow and Tirole [14], Cramton and Stoft [6] and Newbery [15].

2
fixed in the short-run. Moreover, the different technologies in the generation of electricity have
very distinct features. While thermal sources have a typical neoclassical production function with
increasing marginal costs and marketable inputs, renewables rely on non-marketable inputs that
are stochastic and non-storable, but their marginal production costs are negligible.
The paper proceeds as follows. First, we develop a dynamic stochastic model of energy markets
that captures the main features of these markets and provides a tractable framework to analyze the
period-by-period competitive equilibrium. We show that under some assumptions the first welfare
theorem holds. Given a capacity matrix, this model enable us to compute the intertemporal cost
of the energy production. We then calibrate the model. Through calibration, we are able to pin
down the investment cost for unit of capacity of each technology. With the market model and the
investment cost, we are then able to compute the technology mix that minimizes the intertemporal
production-plus-investment cost. This is the optimal energy matrix for a country.
We then use this optimization method together with a constraint to limit thermal production
to a given level. The cost obtained when we compute the constrained optimal matrix with the
amount of thermal generation approaching zero is the cost of complying with the Paris Agreement.
The results enable us to provide policy recommendation for investment in capacity in different
economies. Next, we use the model to simulate what would happen without the Paris Agreement
or any policy intervention, but with a declining price of a unit of intermittent capacity.
Specifically, in our model energy can be produced through hydro, thermal and intermittent
power sources. The behavior of the three different types of power generators is quite different.
While thermal generators produce energy based on marketable inputs - coal, oil, natural gas -
hydro and intermittent sources produce power based on non marketable, renewable and free of
costs inputs, stochastically delivered by nature: sun light, wind, and water. However, contrary to
sunlight and wind, water is storable, and that changes the problem completely. Hydro generators
face a dynamic problem: using water in the present may prevent them from generating energy in
the future. In this economy, the market price correctly signals hydro generators to ration water in
the present for future use.
Our main theoretical result is a version of the first welfare theorem for this economy: that is,
the market equilibrium is Pareto optimal. The main intuition here is the following. First note
that thermal sources and renewables are producing the static optimal every period, while hydro
producers are concerned about their static profit as well as their expected continuation profits. In a
decentralized economy, prices will balance current incentives to produce with incentives to save on
water. If prices were higher then equilibrium prices, hydros would produce more, and if they were
lower, hydros would save more water for future periods of scarcity and consequent higher prices.
This in turn is the same logic that the central planner is using to minimize costs: if water is more
valuable today, then it should use more, and if there is abundant energy being produced, then it
should save on water for future use. This result is important because it shows that a liberalized

3
market is efficient.5 It is also quite useful from a methodological point of view, since it allows us
to focus on competitive equilibria when we perform our welfare and cost analysis.
With these results in hand, we are then able to parameterize the model to data. We chose
to study Argentina, Brazil, and Uruguay. These countries have very distinct natural endowment
of rain and wind and quite different actual capacity matrices. We calibrate and compute the
optimal capacity matrix for these countries. The results show an interior solution for the optimal
matrix for each country, meaning that it is optimal to have a mix of thermal, hydro, and wind/sun
generation. However, these optimal mixes are quite different from one country to another. We are
also able to construct the second-best capacity matrix for each of these countries, which considers
the current matrix already in place. Thus, our model provides policy recommendations for each
of these countries. In particular, while Uruguay seems to be close to the second-best, our model
suggests that Brazil and Argentina should heavily invest in wind generation.
We then proceed to answer the main questions raised in this paper: what is the cost of transi-
tioning to an economy with (close to) zero carbon emission in the countries that we study? And
what happens to emissions if no policy is adopted but investment cost in renewable sources falls?
After recovering the installation costs of the different technologies, we compute the net present
value of such transition. These costs amount to a significant change. In Brazil, for example,6 it
implies a share of renewable almost seven times higher than it is today, which would result in a
total energy cost of almost 50% higher than if the country simply ignores the Paris Agreement.
Argentina faces an even more dramatic challenge: its total costs of transitioning to a low carbon
economy would be almost 90% higher than it is today. In contrast, Uruguay is very close to the
first-best with almost zero emissions.
Finally, we simulate an economy where hydro capacity is continuously replaced by renewable
capacity, something likely to happen if the installation cost of renewable capacity keeps its trend of
falling prices. The numerical results show that when we increase the share of intermittent sources
(i) the profitability of the thermal and hydro plants follow a U-shape relationship, (ii) the mean
price drops due to technology diversification and then rise for high shares of intermittent sources,
(iii) the profitability of the renewable sources drops steadily. These findings corroborate the idea
of a balanced technology mix in long term equilibrium, and has a pessimistic view on the goal of
phasing out thermal plants without significant government policy.

Related Literature
Centralized electricity production has received many criticisms for quite some time. Joskow and
Schmalensee [13] pioneer work pointed to efficiency problems in centralized systems and advocated
for full liberalization and competition. In fact, perhaps the major motivation for the wave of
liberalization that the industry has undergone over the last 25 years was to introduce competition
5
This holds conditional on the installed capacity and assuming away market power.
6
Brazil’s intended Nationally Determined Contribution - iNDC, implies a 43% reduction in carbon emissions by
2030 (see https://www.mma.gov.br/clima/convencao-das-nacoes-unidas/acordo-de-paris).

4
and to move away from centralized decisions.
There is a large literature that analyzes liberalization of power markets in many different
contexts and aspects of it. From the pioneer work of Joskow and Schmalensee [13], to important
contributions such as Green and Newbery [10] and Wilson [17], among many others. Garcia et al
[9] analyzes competition between two hydro generators. Their main concern is market power. Our
paper builds on their setup, and extend their analysis to a continuum of generators in a competitive
environment.
Crampes and Moreaux [5] compute the first best, monopoly and duopoly allocation of a market
with a thermal and a hydro plant. Ambec and Doucet [2] analyze decentralization of a hydroelectric
industry and show that while a monopoly brings market power concerns, a decentralized market
may have suboptimal use of water resources.
Our paper has a fundamental distinction to Crampes and Moreaux [5] and Ambec and Doucet
[2]: our model is not aimed at modelling inefficiencies that may arise from limited storage capacity
or market power.7 Instead, it analyzes how a competitive equilibrium works in a continuum of
generators with distinct technologies and where most generators have zero marginal cost. We show
that even in a competitive market prices never go down to zero.
Joskow and Tirole [14] introduce nonreactive consumers. Ambec and Crampes [1] propose a
model of decarbonizing the electricity generation in the presence of such consumers.
Genc et al. [16] develop a theoretical model of dynamic competition in the electricity generation
sector. They consider a single hydro producer and an oligopoly of thermal generators. Our paper
has some methodological similarities with Genc et al. [16], but their concern is market power while
here we want to compute the cost of changing the energy matrix towards a carbon free matrix,
and the effect of this change on the profitability of the other sources.
There is a growing literature about the energy transition towards a low carbon industry. For a
discussion of the energy transition in Europe, see Fabra et al. [7]. Closer to this paper, Ambec and
Yang [3] relate policies aiming at reducing emissions with electricity trade between countries. Helm
and Mier [12] analyze the efficient adoption of renewable intermittent energy. Fabra and Llobet
[8] study the behavior of renewable technologies with zero marginal costs on electricity auctions.
The paper proceeds as follows. Section 2 lay out the benchmark model with hydro, thermal
and renewable generators, and characterize the competitive equilibrium. Section 3 proves that
the first welfare theorem holds in such a market. Section 4 proves the there is an optimal energy
matrix and it is unique. With these results in hand, we proceed to applications. In section 5, we
calibrate the model for Brazil, Argentina and Uruguai, and compute the optimal energy matrix
for these countries. In section 6, we compute the costs of reducing emissions. In section 7, we
analyze the effect of a substantial entry of renewable sources on the electricity industry reliability.
The last section concludes.
7
Garcia et al. [9] analyze market power in a setting with hydro oligopoly.

5
2 Benchmark Model
2.1 Competitive Equilibrium in a Hydro-Thermal Economy
Consider an infinite horizon model in which there is a unitary inelastic demand for electricity,
which might be supplied by a thermal production source or by hydroelectric power source. We
assume a continuum of identical hydro sources and a single thermal source. Each thermal can
produce an unlimited amount of energy at an increasing and convex marginal cost. The hydro
plants are indexed between [0, θ], where 2 > θ > 1. Each hydro plant can generate energy at zero
marginal cost, but is capacity constrained and atomistic, that is, if a set S ⊂ [0, θ] is producing
´
energy, the energy produced is given by e (S) = i∈S 1di.
At each period t = 1, 2, ..., a hydro generator might have full capacity or empty capacity, which
we will denote by the binary variable {0, 1} . The hydro plant might decide to sell its energy or
not. That is, each hydro plant has a discrete choice and energy is modeled as an indivisible unit.
If it does sell, the plant earns the current market price but will finish the period with empty
reservoirs. If it does not sell energy, it will enter the following period with full capacity. With
these assumptions, we can focus on the extensive margin of the energy production.
In order to capture the aggregate uncertainty that is intrinsic to the production of hydroelectric
energy–but not to the thermal production–, we will assume that at each period t, there are two
possible states of the world. Formally, we assume that with probability π ∈ [0, 1], which is drawn
independently before every period t, the state of nature is ω ∈ {G, B}. Once the state is drawn,
it becomes known to everyone. We eliminate any exogenous idiosyncratic uncertainty and assume
that if the state of the world is G, every hydro generator has full capacity, while if the state of the
world is B, each hydro will have full capacity only if it had full capacity in the previous period
and decided not to sell its energy in that period. Otherwise, its capacity is empty.
The thermal generator has no capacity constraint and faces the same production costs every
period. Its (static) objective function is: maxe π T = pe − c (e), where e is the quantity of energy
sold, c (e) is the cost function and p is the current market price. For convenience, we think of
the thermal source as a price- taker player, capturing the fact that we are actually modeling a
representative generator. The hydro’s static profit function is given by π H = pe. However, recall
that (i) each hydro is capacity constrained and atomistic and (ii) there is a dynamic link, through
the hydro’s reservoir capacity, between the current period’s profit and the expected continuation
payoff starting at the subsequent period.
A public history at time t is denoted by ht and is a sequence of states of the world and energy
sold by each player. The set of all public histories at time t is Ht .8
We consider a Walrasian market in which firms, hydro plants and thermal sources, are price-
8
A behavioral strategy for a hydro generator is defined as:

σH : ∪∞
t=0 Ht → {0, 1} ,

6
takers.9 A competitive equilibrium in our environment is a sequence of prices such that the market
clears in a period-by-period basis, that is, in every history demand meets supply, moreover, the
thermal producer maximizes its myopic profit function and each hydro plant maximizes its dynamic
expected continuation payoff.

Definition 1 (Equilibrium) An equilibrium is a map p : ∪∞ t=0 Ht → R such that the market

clears at every period, the thermal generator is maximizing its per-period payoff and each hydro is
maximizing its expected present value profit.

We will denote hydro plant i0 s expected continuation profit at the start of any period t for which
the state of nature is G by VG . Recall that in state G, every hydro has full capacity, regardless
of what happened in the previous period. In contrast, whenever the state of the world is B, each
hydro’s capacity is dependent on whether it has sold energy in the previous period or not. With
slight abuse of notation, we denote by VB1 the value function of a given hydro when the state is
B, and the hydro has full capacity, while by VB0 we denote the value function of the hydro when
the state is B and it has empty capacity. The value function at state G can be written as the
maximum between selling and not selling the energy:

VG = max pG + δ πVG + (1 − π) VB0 , δ πVG + (1 − π) VB1 .




(1)

Regardless of the state of the world, whenever a hydro sells its energy at price p, it empties
its reservoir, which means that it will only be able to sell energy again whenever the state of the
world becomes G again. Thus, we may write the payoff VB0 for any period in which the state is B
and the hydro has an empty reservoir as:

VB0 = δ πVG + (1 − π) VB0 .

(2)

The value function at B when the capacity is full is given by:

VB1 = max pB + δ πVG + (1 − π) VB0 , δ πVG + (1 − π) VB1 .




(3)

Before we proceed let us define p∗ = c0 (1), that is, p∗ is the Walrasian price when only the
thermal generator supplies the market.
In order to better understand the mechanics of our model, let us first consider the case in which
the number of hydro plants selling in a particular given period exceeds the demand. Given that
and for the representative thermal generator it is:

σT : ∪∞
t=0 Ht → R.

9
There are over a thousand hydro plants in Brazil, which is our prototypical example.

7
 Figure 1: Only hydro production (p=0)

Figure 2: Equilibrium (t=1): good state

each hydro supplies at a zero marginal cost, the Walrasian price in that period would be zero, as
shown in Figure 1.
In contrast, if the number of hydro sources selling at a particular period is smaller than the
unitary demand, then thermal sources are used to clear the market. Such a scenario is presented
in Figures 2 and 3 below.
The market price for such a period in which hydro sources do not supply the entire demand is
given by the marginal cost of the thermal production at the market-clearing quantity.
In equilibrium, hydro plants must be maximizing their infinite stream of profits discounted by
the discount rate δ < 1. For that reason, a price of zero as presented in Figure 1 will not happen
in equilibrium, regardless of the history. This is a consequence of the fact that hydro plants are
forward-looking and subsequent periods have a positive probability of some water-scarcity. Hydro

8
 Figure 3: Equilibrium (t=2): bad state

plants anticipate a positive price in the future and thus prefer to keep water in their reservoir to
profit when price is high. Thus, a first implication of our model is that prices are always positive,
despite the fact that at given periods there might be excess supply of zero-marginal-cost suppliers.
Figures 2 and 3 show the prices in a good state and in a bad state, respectively. Note that
prices are higher in bad states. That is the intuition of the equilibrium price dynamics in this
market.
We are now ready to prove our first two results: first, we prove that an equilibrium always exists
in these markets, then we provide a characterization of an equilibrium, that is, a few necessary
conditions for equilibrium.

Proposition 1 (Existence) An equilibrium always exists in this market.

Proof. The proof is by construction and we provide it in the Appendix.

Lemma 1 (Equilibrium Properties) An equilibrium must satisfy the following properties:

i) Prices are weakly lower than the price given by the production of thermal sources exclusively
∀h, ph ≤ p∗ .
ii) If the price is the maximal one, p = p∗ , then firms have strict incentive to sell energy.
iii) If at period t firms have strict incentive to sell, then at period t + 1 if the state is B : p = p∗ .
iv) For any period t, firms weakly prefer to sell energy versus not selling.
v) In equilibrium, at state G, firms are indifferent between selling and not selling energy.

The competitive price is given by the marginal cost of the marginal producer. Recall that in
the good state G, where all hydro plants have water in their reservoirs, it is not an equilibrium
to have only hydro generation. If this was the case, p = 0, and the hydro generators would have
an incentive to save water for the next period. An important implication of this is that there is
always thermal generation, qT > 0, which gives the next lemma.

9
Lemma 2 (Thermal is the marginal producer) In equilibrium, pt = C(qTt ).

The next lemma characterizes the market price dynamics in the competitive equilibrium.10

Lemma 3 (Price Dynamics) Let pG be the price in the good state and pB be the price in a bad
state with pB < δ (1 − π)p∗ . Then, in equilibrium we have

pG = δ (1 − π) p0B (4)

and
pB ≥ δ (1 − π) p0BB (5)

Equation 4 shows the relationship between the price in the good state, pG , and the price in the
bad state, pB . It shows that the price in the good state is the expected discounted value of the
price in the bad state next period. It comes from the fact that hydro generators in the good state
are indifferent between selling or not energy. Similarly, equation 5 shows the relationship between
the price in a bad state following a good state, pB , and the price in a bad state following a bad
state, pBB . The inequality comes from the fact that hydro generators may strictly prefer to sell
energy in this situation.

Remark 1 (Price Cap) To finalize this section, let us consider an important regulatory tool:
price caps. Assume that the price of energy is always constrained to be pt ≤ p̄ < p∗ . Our next
result follows immediately from Lemma 1 (ii).

Lemma 4 If pt = p̄, then all hydro plants sell their capacity in that period.

Now, consider a history in which there is a mass µ of hydro plants with full capacity, and let
µ < 1. In any such case, if the expected future price is not high enough, the price of energy might
be p∗ (or p̄ in the case of a binding price cap) in the subsequent period. If, in equilibrium, all hydro
plants find it profitable to sell their currently stored reservoirs, that is, µ is such that e = µ, then
we say that we are in an extreme scenario.
Using Lemma 4 we have that under a binding price cap, hydro plants sell weakly more energy
every period than when there is no price cap. This implies that under a price cap, the economy
has a higher probability of reaching extreme scenarios. We state this result below.11

Proposition 2 Assume that a price cap has been imposed: p̄ < p∗ . Denote this economy by Ec and
construct an identical economy, but with no price cap, denoted by Ef . Then, whenever the state is
G, the energy produced by hydro sources is higher under Ec than it is under Ef , ec > ef moreover,
Ec has a higher probability of reaching extreme scenarios than f.
10
Whenever the water storage is low enough so that the current price using all water left in storage is high enough
(namely, when p > δ (1 − π)p∗ ) the hydro sources with water in storage will strictly prefer to sell their water.
11
This result is simlar to the one obtained by Garcia et al. (2001), but theirs is a duopoly of hydro generators,
while we have a competitive market with a mix of technologies.

10
2.2 Intermittent Sources
In our benchmark model, we deliberately ignored intermittent power sources. By intermittent
sources, we consider all power plants that satisfy two conditions (i) a zero marginal cost of pro-
ducing energy and (ii) lack of capacity to store energy. Note that the hydro sources satisfy (i), but
they can store energy in their reservoirs.
Intermittent sources include most notably wind power, solar energy and run-of-river plants
(plants that generate hydro energy, but have no reservoirs). For environmental reasons, this
source of power is generating great interest throughout the world. It is already quite common in
some places: Denmark, for example, generates around 40% of its electricity from wind. China,
as another example, has recently announced ambitious plans to increase wind and solar power
capacity. In Brazil, run-of-river plants are common and relatively representative. We will model
all of these different sources as a continuum of producers with zero marginal cost and no reservoirs.
In an electricity system with thermal and hydro sources, the entry of intermittent generators
has two effects: it decrease the ratio of storage to total capacity, and it diversifies the stochastic
processes that generate the input.
In order to have a better understanding of the role of intermittent sources, we will try to
disentangle these two effects by analyzing two cases. First, intermittent generators with no storage
but the same stochastic process as the hydro sources. We will think of these sources as being run-
of-river plants, so that the uncertainty inherent in the production is the same as with the hydro
generators. This is useful as a benchmark, but it also provides us with policy recommendations
for investment in this type of generator. Second, we study the, perhaps, more interesting case in
which intermittent sources follow a stochastic process that is distinct from that of hydro sources.
As an example, wind and rain may have distinct and independent probabilities of occurring. For
the first case, we provide an analytical solution, whereas for the second case, the complexity of the
problem is substantially increased and, therefore, we solve the model numerically in section 5.

2.2.1 Run-of-River

In order to understand the role that intermittent sources play, lets consider the case where we
replace hydro plants with storage by run-of-river plants without it. Assuming that intermittent
sources have mass η < 1, it means that η + θ is constant throughout the analysis. It is perhaps
not surprising that energy should be weakly cheaper in G when all reservoirs are full and the
intermittent plants can generate energy.
What is more interesting is that there is a critical level η, such that for any η ≤ η, the market
equilibrium is the same as the one without intermittent sources. Also, this critical capacity level
is equal to the amount of energy that is produced by the hydro plants in the state G when there
is no intermittent capacity installed.

Proposition 3 (Intermittent Power Plants: Threshold capacity for run-of-river plants)

11
i) There is a critical level of intermittent capacity η, such that for any η ≤ η, the market equilibrium
is unchanged with respect to the benchmark case.
ii) Let eh be the energy generated by the hydro plants in the benchmark case. Then we have,
η = eh .

The intuition behind this result is the following. The energy from the intermittent sources will
be used in the good state anyway. If their capacity is less than or equal to eh , the energy produced
on state G is still eh , so that the share of hydro that saves water to the next period is θ − eh , as
in the benchmark case.
For cases where η > η, the equilibrium differs. In order to understand the effects of intermittent
sources, consider two economies Ei and Ej with a mass of intermittent sources lower in Ei : ηi <
ηj . To have a comparable situation, we assume that the total installed capacity of hydro plus
intermittent is the same in the two economies, and only the share of each technology differs. That
is,
θi + ηi = θj + ηj , (6)

and hence, θi > θj . Also, for the next proposition assume that at least ηj > eh , or both ηi and ηj
are greater.

Proposition 4 (Intermittent Power Plants: Higher price volatility) Consider two economies
Ei and Ej with a mass of intermittent sources lower in Ei : ηi < ηj , where ηj > η. Also assume that
θi + ηi = θj + ηj . Then we have the following:
i) The equilibrium prices are lower in Ej in periods G: pjG < piG .
ii) The equilibrium prices are higher in Ej in periods B: pjG > piG .
iii) Price volatility is higher in Ej .

This result has one important policy implication. Competitive energy prices in the good state
will be lower with more intermittent sources operating. With no interference, this lower price will
signal intermittent sources not to enter the market until perhaps demand increases and it becomes
profitable for these sources to enter again.
A subsidy is a common policy for intermittent and environmental friendly sources. It induces
entry of intermittent sources, and it reduces energy prices, as shown. These lower prices requires
further incentives to induce further entry of clean generators.
Up to this point, the stochastic process governing the delivery of hydro and intermittent energy
is identical. It simplifies the problem greatly, but it is not realistic. We will relax this assumption
in section 5, when we will analyze the effect of intermittent sources through a numerical simulation
of the social planner’s problem. But to do this, we need to guarantee that the first welfare theorem
holds, meaning that the solution to the planner’s problem gives rise to the same allocation as that
of a market economy. The next section shows that the first welfare holds.

12
3 Social Planner
The social planner chooses a production plan to minimize the intertemporal sum of discounted
costs.
∞
X
Wt = δ t Ct (Q) , (7)
t=0

in which Ct (Q) represents the total cost of producing Q units of energy.

We look at the planner’s problem considering a capacity θ of hydro and R of renewable sources.
We will treat both θ and R as state variables since this will be essential to study the long-run
optimal policy, that is, when the planner is allowed to choose the optimal energy matrix. We will
do this in section 4. In this section we will focus on the case where the matrix is given. This is
the short-run problem: for a given matrix, what is the optimal policy of the planner?
We assume that there is a constant inelastic demand D every period, and the cost structure
is the following: the hydro plant has zero marginal cost and the thermal plant’s cost function is
C(qT ), with C 0 ≥ 0 and C 00 > 0. We further assume that C 0 (0) = 0. The thermal plant has enough
capacity to supply the entire demand at a cost C(D). We restrict R to the interval [0, D] and θ
to the interval [0, θ∗ ]. Denote by CG the cost of production when the state is good for renewables
and CB otherwise. The planner’s problem in the recursive formulation is :

Ci (X, θ, R) = min {C(qT ) + δ (πγCG (θ, θ, R) + (1 − π)γCG (X − qH , θ, R) + π(1 − γ)CB (θ, θ, R)) +
qT ,qH ,qR

(1 − π)(1 − γ)CB (X − qH , θ, R)}

s.t
D = qT + qH + qr
qH ≤ min{X, D}
qR ≤ R̄

u R, if i = G
h
R̄ =
u R, if i = B
l

X − q , with probability 1 − π
H
X0 =
θ, with probability π

Where qT , qH , and qr stand for quantity produced of thermal, hydro and renewable energy,
respectively. X and X 0 denote the total amount of water stored in the reservoirs in the current
and subsequent period.
The first restriction says that hydro plus thermal plus intermittent must match the total de-

13
mand, the second states that hydro generation is constrained by the amount of water available
and cannot exceed the demand, the third states that the production of intermittent sources cannot
exceed the available capacity R̄, the fourth states that the capacity of producing from intermittent
sources in a period is given by a factor uh or ul (with uh > ul ) multiplied by the total capacity
R.12 Finally, the last equation shows the law of motion of the state variable X.
The total cost function is defined as the sum of thermal plus hydro plus intermittent generation:
C = CT (qT ) + CH (qH ) + Cr (qr ) . Since the costs of the hydro and of intermittent sources are zero,
we write the one period cost function as C = C (qT ).
Our next result is straightforward and comes directly from the fact that there is no storage of
renewable energy in our model.

Lemma 5 (Full use of renewables) The planer will always choose qR = R̄.

The next result states that the planner will always choose to employ some thermal sources in the
energy production. Before we proceed with the proofs, note that that the cost function C(·) is
decreasing in the first argument. The following result shows that it can never be optimal to use
no thermal in the current period and some thermal sources on the following period. This result
comes from the convexity of the cost function. We prove this result by showing that such a policy
has a profitable deviation in which the planner uses some thermal source to produce in the current
period and saves water for the future.
The next proposition shows that the planner will always use some thermal energy sources. The
proof is on the Appendix.

Proposition 5 (Always optimal to use thermal sources) qH = D − R̄ is never an optimal

choice. (As a consequence, thermal is always produced).

From now on, it will be convenient to consider the equivalent problem of maximizing the
function −C(qt ) instead of working with the minimization problem above. Then we have that
−C 0 (qt ) < 0 and −C 00 (qt ) < 0, so the objective function becomes a strictly concave function. Next
we write the problem in terms of choosing the remaining hydro capacity X 0 = X − qH . Thus, we
can rewrite the above problem as:

V i (X, θ, R) = max
0
{−C(D − R̄ − X + X 0 ) + δ(πγV G (θ, θ, R) + (1 − π)γV G (X 0 , θ, R)+
X

π(1 − γ)V (θ, θ, R)) + (1 − π)(1 − γ)V B (X 0 , θ, R)}

B
s.t
max{0, X + R̄ − D} ≤ X 0 ≤ X

u R, if i = G
h
R̄ =
u R, if i = B
l

12
For the proofs we use uh =1 and ul =0. For the calibration excercise, we use more realistic values.

14
 Where the maximum is there to ensure that the energy production never exceed the demand.
We are now able to state our main theoretical result, which combines the results in this section
with the ones of section 2.1. In the Appendix, the proof provides some properties of the value
function, which are necessary for the theorem.

Theorem 1 (First Welfare Theorem) The competitive equilibrium is Pareto-optimal.

Proof. See appendix.

In section 5 we will use the first welfare theorem to implement our numerical analysis. Before
we proceed, let us discuss about a very important issue in markets with different technologies.

4 Optimal capacity
We assume that there is an installation cost Ch per unit of hydro capacity and CR per unit of
renewable capacity. Our objective is to show that there is an optimal energy matrix.
We solve this problem from an ex-ante perspective. We assume that the planner can choose
a capacity in the space E := [0, θ̄] × [0, R̄], with θ̄ large enough13 and R̄ ≤ D, then the planner’s
problem of choosing the optimal energy matrix will be:

max V (X, θ, R) − Ch θ − CR R
(θ,R)∈E

Theorem 2 (Optimal Capacity Matrix) There exists a unique optimal capacity matrix (θ∗ , R∗ ) ∈
E.

Proof. Note that this problem is well defined since V G (·) is continuous and bounded and E is
compact, so the existence of (θ∗ , R∗ ) ∈ E follows directly. Uniqueness comes from the fact that V g
is strictly concave.
Note that this optimal capacity will depend on the combination of the parameters of the model,
which renders an analytic solution infeasible. We can, however, solve this numerically given the
parameters.
To illustrate how the optimal capacity relates to the installation costs, we present the following
three figures. In figure 4, we show the relation of the planner’s value function with the capacity
installed of hydro and intermittent sources. The main idea to grasp from this figure is that the
value function is (obviously) increasing in the installed capacity, but it is also clearly concave. We
prove this in Lemma 7.

In Figures 5 and 6, we present (i) the optimal hydro capacity as a function of the installation
costs and (ii) the optimal capacity of intermittent sources as a function of these costs. The
important message here is that the optimal matrix will be interior for reasonable cost parameters.
13
Precisely, we need only that θ̄ ≥ θ∗ , where θ∗ is the optimal capacity in the relaxed problem.

15
 Figure 4:
Value Function and Capacity

-0.5
-1
-1.5
-2
-2.5
-3
-3.5
-4
-4.5

1
0.8 3
Re 2.5
ne 0.6 2
wa
ble 0.4 1.5
Ca 1 ity
pa
city 0.2 capac
0.5 Hydro
0 0

Figure 5:
Optimal hydro given Costs

0
4
3
Re 4
n ew 2 3
ab
le 1 2
C os Cost
t 0
1 Hydro
0

Second best

For a fixed initial matrix (θ, R), with θ < θ̄ and R < R̄. we can ask what is the optimal investment
plan. Define now the capacity expansion space as T := [0, θ̄ − θ] × [0, R̄ − R], the problem is now:

max V G (θ + ∆θ, θ + ∆θ, R + ∆R) − Ch ∆θ − CR ∆R

(∆θ,∆R)∈T

Again uniqueness comes from the strict concavity of the value function.

Proposition 6 (Second Best Capacity Matrix) There exists a unique second best capacity
capacity (θ2∗ , R2∗ ) ∈ E.

We use these results, together with a calibration of the parameters of the model, to compute

16
 Figure 6:
Optimal renewable given Costs

0.5

0 4
4
3
3
Re 2
new 2
ab
le 1
Co 1 Cost
st
0
Hydro
0

the optimal matrix, the second best conditional on the current capacity matrix and the matrix
aimed by the Paris Agreement, without thermal generation.

5 Quantitative Analysis
We develop a strategy to solve the model numerically. We parameterize the model with data
from Brazil, Uruguay and Argentina. These three countries are representative in that they differ
significantly in terms of their natural characteristics and in their current energy matrix. We obtain
(i) the unconstrained optimal capacity matrix for each of these countries and (ii) the second best
optimal matrix, which has an important policy implication: how should these countries invest
given their current capacity matrix.

5.1 Parameterization
Overall strategy

We now describe our calibration strategy. We use one year of data with monthly frequency.
Below, we describe our algorithm to retrieve the parameters in that sequence. In order to get the
investment cost we assume free entry in the generation market, so that we equate the expected
profits with the investment cost. We assume a unitary demand and a quadratic cost function.

• First we choose δ in order to match a target real interest rate r, considering a monthly

12
frequency: 1δ = r.

• We set γ = 0.5. We then sort the monthly observations on the wind capacity factor and
choose uh as the average of the top 6 months and ul as the average of the bottom 6. This
way, the average capacity factor in the model coincides with the one observed in the data.

17
 • R is chosen in order to match the average share of wind energy production in the total energy
generation, which is given analytically by sr = R(γuh + (1 − γ)ul ).
θ
• We set θ to match the observed relation between hydro and wind installed capacities: ω = R

• We find π in order to match the average share of hydro energy production in the total energy
generation, this cannot be done in a closed form, so we do this procedure numerically.14

• Given the above calibration we simulate the model 10000 times to get the expected profit
for hydro and wind energy sources, Ph and Pr respectively. We then find the cost of each
source by equating profits and investment cost:

Ph = ch θ and Pr = cr R

Tables 1, 2 and 3 summarize the parameters for Brazil, Uruguay and Argentina used in the
quantitative analysis.

Table 1: Parameter Values: Brazil

Brazil 2018
Parameter Value Target
Discount factor δ 0.997 Gross interest rate of 1.036(yearly basis)
Wind scaling factor uh 52.09% average wind capacity factor of 41.91%
Wind scaling factor ul 31.72% average wind capacity factor of 41.91%
Intermittent capacity R 0.2098 8.79% share of wind in energy production
Hydro Capacity θ 1.4851 theta/R = 7.08
Probability of rain π 73.92% 74.66% share of hydro in energy production
Installation cost of Hydro ch 17.11 Expected profits equal investment costs
Installation cost of Intermittent cr 23.2636 Expected profits equal investment costs

5.2 Main Results

In Table 4 we report the optimal capacity matrix and the second best optimal matrix. The former is
the unconstrained optimal choice of capacity given the parameters of the economy, but disregarding
the actual capacity. The latter represents the optimal capacity matrix given the parameters of the
economy but also the current matrix.
Our policy recommendations are based on the second best results, since we do not expect
capacity do be removed. The main suggestions are: (i) Brazil should invest in wind energy, (ii)
14
We use the Matlab routine fzero to find a zero of a residual function between the average share generated by
the model (for a given π) and the target. The routine brackets the value of π in an interval [a, b] and progressively
shrinks the bracket until the distance between the extremes is small.

18
 Table 2: Parameter Values: Uruguay
Uruguai 2017
Parameter Value Target
Discount factor δ 0.997 Gross interest rate of 1.036(yearly basis)
Wind scaling factor uh 32.01% average wind capacity factor of 28.69%
Wind scaling factor ul 25.38% average wind capacity factor of 28.69%
Intermittent Capacity R 1.0946 31.41% share of wind in energy production
Hydro Capacity θ 1.1144 theta/R = 1.01
Probability of rain π 81.33% 59.12% share of hydro in energy production
Installation cost of Hydro ch 11.60 Expected profits equal investment costs
Installation cost of Intermittent cr 8.86 Expected profits equal investment costs

Table 3: Parameter Values: Argentina

Argentina 2017
Parameter Value Target
Discount factor δ 0.997 Gross interest rate of 1.036(yearly basis)
Wind scaling factor uh 100% average wind capacity factor of 34.86%
Wind scaling factor ul 0% average wind capacity factor of 34.86%
Intermittent Capacity R 0.0129 0.45% share of wind in energy production
Hydro Capacity θ 0.6614 theta/R = 51.24
Probability of rain π 0.4800 28.96% share of hydro in energy production
Installation cost of Hydro ch 73.07 Expected profits equal investment costs
Installation cost of Intermittent cr 63.24 Expected profits equal investment costs

19
Uruguay is already close to the second best (but has experienced over investment in wind capacity
and under investment in Hydro sources) and (iii) Argentina needs to massively invest in wind
generation. The percentage changes are reported on last column of Table 4.
If we consider the first best, (i) Uruguay has over invested in wind generation, (ii) Brazil has
over invested in hydro sources, and (iii) Argentina in hydro plants.
Table 5 shows the optimal share in terms of quantity generated. Note that the very large
percentual increase in Argentina’s wind installed capacity, corresponds to a 9% increase in the
total amount generated by this technology.

Table 4: Optimal Capacity Matrix

Brazil 2018
Source Actual Capacity Optimal Capacity % change Second Best % change
Hydro 1.4851 1.4293 -3.75% 1.4851 0%
Wind 0.2098 0.3055 +45.63% 0.2658 +26.72%
Uruguay 2017
Source Actual Capacity Optimal Capacity % change Second Best % change
Hydro 1.1144 1.2012 +7.79% 1.1320 +1.57%
Wind 1.0946 0.9518 -13.05% 1.0946 0%
Argentina 2017
Source Actual Capacity Optimal Capacity % change Second Best % change
Hydro 0.6614 0.5141 -22.28% 0.6614 0
Wind 0.0129 0.2738 +2020% 0.2147 +1563%

20
 Table 5: Optimal share of generation by technology

Brazil 2018
Source Actual share Optimal share Second Best share
Hydro 74.66% 71.43% 73.57%
Wind 8.79% 12.80% 11.14%
Uruguay 2017
Source Actual share Optimal share Second Best share
Hydro 59.12% 63.46% 60.08%
Wind 31.41% 27.31% 31.41%
Argentina 2017
Source Actual share Optimal share Second Best share
Hydro 28.96% 23.23% 28.54%
Wind 0.45% 9.54% 7.49%

6 The Cost of Decarbonizing the Energy Matrix

In this section we simulate for Brazil, Uruguay and Argentina the cost of phasing out thermal
generation and carbon emission. Technically, we compute the constrained optimal capacity matrix
and the production-plus-investment cost associated with it. Figures 7, 8 and 9 show the impact
of reducing thermal generation from current levels towards zero. The vertical axe shows the
normalized cost, with the current cost normalized to one. As one moves towards lower levels of
thermal generation the cost increases.
Graph 7 shows that Argentina’s current thermal generation accounts for more than 60% of
total generation. To cease thermal generation, total electricity cost would increase by almost 90%.
The results are less dramatic for Brazil and Uruguay, figures 8 and 9, respectively, since these
countries already rely mostly on clean energy sources. For Brazil, it will imply in a 40% increase
in energy costs, while for Uruguay it will be a 20% increase.
The conclusion is that (i) the costs are substantial, and in some cases, perhaps too large. And
(ii) there is a strong heterogeneity in the distribution of the transition cost across the countries,
which is a function of the current energy matrix and the endowment of energy sources, like rivers,
wind and sunshine.
Interestingly, the model predicts that thermal generation will be mostly replaced by wind
capacity. We first thought that it was caused by the installation cost, but it is not true since the
installation cost of wind capacity is higher than that of hydro in Brazil. The explanation is the
following. There is always some wind generation in the data set we use for calibration. It happens
since wind turbines are spread over several distant areas within the three countries, with distinct

21
Figure 7:

Figure 8:

22
wind patterns. Intuitively, there is always wind somewhere.
The same is not true for raining; it only rains on some periods, with zero rain on several periods
we observe. Despite the storage capacity of hydro generators, there is always a probability of a
sufficient long sequence of dry periods which would lead to insufficient electricity supply. On the
other hand, intermittent wind generators can always generate electricity. For this reason, and
surprisingly, the model optimally increases wind generation capacity as a way to avoid shortages.
Figure 10 shows that Argentina will need 140 times more wind generation capacity than it
has today. It is a large but reasonable number since Argentina has only 1% of wind plants. In
Brazil, wind capacity needs to increase almost seven times to account for the decrease in thermal
generation, as shown in figure 11.
Figure 12 illustrate the fact that Uruguay already has a large base of wind generators, such that
it only needs to double its wind capacity to compensate for the reduction in thermal generation.

23
Figure 9:

Figure 10:

24
Figure 11:

Figure 12:

25
7 Extension: Intermittent Power Sources
Contrary to the previous section, where we computed the optimal capacity matrix, in this section
we analyze what will happen to thermal generation if we leave it to the market, without any en-
vironmental policy. One natural question is to ask: what would happen to electricity generation
without the Paris Agreement? To try to answer it, we compute the profitability of the thermal
plants when we increase the share of renewable capacity on the energy matrix. The cost of re-
newable technology has fallen substantially over the last decade, which increases their profitability
and rate of adoption.
One concern about the dominance of intermittent renewables is that by decreasing the average
electricity price it might expel other more reliable technologies but with higher costs, such as
thermal producers. This would, as the argument goes, reduce the reliability of the system due to
intermittency.
In order to analyze whether it is true that the entrance of intermittent sources continually
decreases the expected profitability of thermal plants, we simulate a system using the parameters
calibrated to Brazil from table 1 and we look at the profitability of thermal plants as a function
of the fraction of intermittent sources in the economy. In order to have a fair comparison between
different industry structures, we assumed that increases in renewable’s capacity replaces hydro
plants, so that total generation capacity remains constant.
Figure 13 shows the profitability of the thermal plant against the share of renewables installed.
Note that the profitability of the thermal plants follows an U-shape relationship with the amount
of intermittent renewable installed.
Two forces are at play in this case. First, when wind capacity start replacing hydro capacity
electricity prices drop on average, as shown in figure 14. It happens because wind and rain follow
distinct stochastic processes, and now there is a higher probability that in any period there is either
rain or wind or both, lowering electricity cost. This effect dominates at first, and it highlights the
gain from a more diverse portfolio of generating technologies.
But as we decrease the share of hydro in favor of intermittent renewable capacity, the hydro
plants lose the ability to transfer water across time, saving water for scarcity periods. It increases
the probability that the thermal plant find itself in a situation where it has to produce a large
quantity at a high price. This second effect dominates the first for higher shares of renewables.
Electricity average price drops steadily until a threshold R∗ . For this point on, the loss of
storage capacity outweights the gains from diversification. Perhaps unexpectedly, figure 15 shows
that price volatility drops until a threshold R∗∗ , then raises sharply. Our first thought was that
price volatility would only increase as we decrease the share of hydro plants with storage capacity,
since they are the ones which act to smooth the price. Again, the diversification effect is stronger
than the smoothing for lower levels of renewable capacity.
The profitability of the wind plants per unit of capacity also follows a U shaped curve.

26
Figure 13: Thermal technology profitability and share of renewables
Total thermal profit
11

10

Thermal Profit
8

4
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
R

Figure 14: Average prices and share of renewables

Average price
0.19

0.18

0.17

0.16
Average price

0.15

0.14

0.13

0.12

0.11

0.1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
R

Figure 15: Price volatility and share of renewables

price standard deviation
0.2

0.19

0.18

0.17
standard deviation

0.16

0.15

0.14

0.13

0.12

0.11
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
R

27
 It suggests two long term conclusions. First, thermal generation will not be driven out of the
industry by market forces. High shares of intermittent generation will increase the profitability of
the thermal plants. This results corroborate the idea of a ’clean energy paradox’, where the entry
of renewable sources become progressively harder as its participation on the market increases.
Second, the generation technology mix tends to be balanced, in the sense that it is unlikely that
renewable generation technologies with its zero marginal costs drive other technologies out of
the market. Public policy is mandatory in order to achieve a corner solution with zero thermal
generation.

8 Conclusion
This paper characterizes the competitive equilibrium of an electricity industry comprised of ther-
mal, hydro and intermittent sources, and show that this equilibrium is Pareto optimal.
The price is determined by the marginal generator, usually a thermal plant. The fact that
water is storable introduces dynamics in the problem. Even with zero cost, hydro generators do
not sell at zero price. An important result shows that the market price signals water scarcity in
the future and induce hydro generators to refrain from producing, even if they have lower marginal
cost than thermal generators.
Given capacity investment cost, we use the model to show that there is an unique optimal energy
matrix. We then calibrate it with data from Argentina, Brazil and Uruguay. The calibration allows
us to obtain the investment cost for each technology in each country, and then use the model to
compute the optimal energy matrix for each country. The results show in which technology it is
optimal for these countries to invest. We obtain that Argentina and Brazil should invest in wind
generation, while Uruguay should invest in hydro sources.
We then use this method to compute the cost of complying with the Paris Agreement and
to cease thermal generation. The results show a large but heterogeneous cost to adapt to the
constraints imposed. Argentina will have a roughly 90% increase on the cost of its energy matrix.
Brazil comes next with a 40% increase and last comes Uruguay with only 20% increase.
The model prescripts that the thermal generation should be replaced by intermittent wind
plants. As surprisingly as it is at first, the intuition of the result makes sense. According to the
data we use, there is always wind generation. There are different wind intensities withing an year,
but never zero generation. It implies that wind intermittency can be compensated by a large
wind capacity. On the other hand, hydro generation has storage but also a positive probability of
observing a sufficient long sequence of dry periods.
The heterogeneity on the cost of complying with the Paris Agreement comes from the current
amount of thermal generation. Argentina has a high share, more than 60%, and Uruguay has a low
share, less than 9%. It also depends on the current share and installation cost of wind generators.
Next, we use the model to simulate what would happen to the energy matrix without the Paris

28
Agreement. We conclude that the observed trend of falling prices of intermittent technologies is not
enough to shut down carbon emissions from electricity generation. There is a U-shaped relationship
between the profitability of thermal plants and the share of intermittent sources. For high levels of
intermittent sources, the thermal plants will produce large amounts when the intermittent plants
are not producing and prices are high.
However, the introduction of batteries and storage of energy generated from renewable sources
may change this scenario. In this case, the thermal plants may indeed become obsolete and perhaps
end up exiting the electricity market.

Appendix A
Proof of Proposition 1. We provide an algorithm to compute the equilibrium:
The first equation comes from VB0 = δ (πVG + (1 − π) VB0 ) :

δπ
VB0 = VG . (8)
1 − δ (1 − π)
The second equation comes from VG = pG + δ (πVG + (1 − π) VB0 ), which leads us to:

1 − δ (1 − π)
VG = pG . (9)
1−δ

Lemma 5 (indifference at G) leads us to:

1


VG = δ πVG + (1 − π) V1,B , (10)

where
1
= p1 + δ πVG + (1 − π) VB0 .


V1,B (11)


We have four equations and five unknowns VG , VB0 , V1,B1
, pG , p1 . These four equations hold for
any equilibrium.
Precisely, we will work with the following steps to find the equilibrium:
Step 1) Assume that V1,B 1
> δ (πVG + (1 − π) v̄), where v̄ ≡ p∗ + δ (πVG + (1 − π) VB0 ).
Step 2) Compute the share of hydro plants with full capacity in state (1, B) .This share is
θ − 1 + T, where T is the share of energy produced by a thermal generator. How is T computed?
pG = c0 (T ). Denote this quantity T by TpG .
Step 3) Compute p1 . This price is given by: p1 = c0 (2 − θ − T ), where the number in paren-
thesis is calculated from the supply of thermal that is needed after every hydro with full capacity
supplies.
Step 4) We now have p1 , solve the system of four equations and four unknowns above.
1
Step 5) Check if step 1 is correct, that is, if V1,B > δ (πVG + (1 − π) v̄). If yes, then the

29
equilibrium for the dynamic game has been solved. If no, then from lemma 4 add the following
1 1

1
two equations: V1,B = δ πVG + (1 − π) V2,B and V2,B = p2 + δ (πVG + (1 − π) VB0 ) .
1
We now have 6 equations and 7 unknowns. Assume that V2,B > δ (πVG + (1 − π) v̄) and proceed
as before:
Step 2.1) Compute the share of hydro plants with full capacity in state (1, B) .This share is
θ − 1 + TG , where TG is the share of energy produced by a thermal generator in state G. How is
TG computed? pG = c0 (TG ).
Step 3.1) Compute p1 . This price is given by: p1 = c0 (2 − θ − TG ), where the number in
parenthesis is calculated from the supply of thermal that is needed after every hydro with full
capacity supplies.
Step 4.1) We now have p1 , solve the system of four equations and four unknowns above.
1
Step 5.1) Check if step 1 is correct, that is, if V1,B > δ (πVG + (1 − π) v̄). If yes, then the
equilibrium for the dynamic game has been solved. If no, then from lemma 4 add the following
1 1

1
two equations: V1,B = δ πVG + (1 − π) V2,B and V2,B = p2 + δ (πVG + (1 − π) VB0 ) .

Proof of Lemma 1. i) p ≤ p∗ . Given that we have imposed that the thermal generator is a
myopic optimizer, p∗ is the maximum possible static price in this dynamic environment.
ii) If p = p∗ firms have strict incentive to sell energy. If the firm sells, it cashes in the maximum
price possible. Its payoff is: p∗ + δ (πVG + (1 − π) VB0 ). If it does not sell, it has a zero payoff in
the current period and in all future periods that it does not sell. If it sells only t periods ahead, it
will get a payoff δ t p + δ t+1 (πVG + (1 − π) VB0 ) < p∗ + δ (πVG + (1 − π) VB0 ).
iii) At t all firms will sell and empty their reservoirs. Thus, all firms will either enter the
following period with full reservoirs (G) or empty reservoir (B). In the latter case, only the
∗
thermal generator will supply energy and the price will be p . 
0 1
iv) We need to show that: pt + δ (πVG + (1 − π) VB ) ≥ δ πVG + (1 − π) V(h,B) . Suppose that
there exists some period t and some history ht for which selling is strictly worse than not selling
energy. This means that nobody sells and p = p∗ . By lemma 2 we have a contradiction.
v) Suppose that selling is better than not selling at some period t and state G. Then, all firms
will sell and price must be 0. (Recall that the mass of firms exceeds the demand and they have
zero marginal cost). If the next period is B, the mass of hydro plants withe full capacity is θ − 1,
which, by assumption is smaller than 1, that is θ − 1 < 1 and thus, price must be p > 0, since
thermal production would be needed to clear the spot market. Thus, the payoff from selling would
be: 0 + δ (πVG + (1 − π) VB0 ) which is small than waiting a period and with probability π being
in the same situation as having sold, but with probability 1 − π having the possibility of selling
energy at a positive price.

Proof of Lemma 3. We repeat the Bellman equations of the hydro firm in the three relevant
states: good (VG ), bad with water (VB1 ) and bad without water (VB0 ):

30
 VG = max pG + δ πVG + (1 − π) VB0 , δ πVG + (1 − π) VB1 ,


VB0 = δ πVG + (1 − π) VB0 ,

VB1 = max pB + δ πVG + (1 − π) VB0 , δ πVG + (1 − π) VB1 .



In equilibrium, on the good state generators need to be indifferent between selling and not
selling energy:
pG + δ πVG + (1 − π) VB0 = δ πVG + (1 − π) VB1 ,

what simplifies to
pG = δ (1 − π) VB1 − VB0 .


(12)

Applying the same reasoning to generators with water on the bad state we get

pB = δ (1 − π) VB1 − VB0 .


(13)

Using the definitions of VB1 and VB0 , we have

VB1 − VB0 = pB + δ πVG + (1 − π) VB0 − δ πVG + (1 − π) VB1 = pB .

(14)

Substituting equation (14) in equations (12) and (13), and letting p0B be the price on the bad
state next period, we have

pG = δ (1 − π) p0B (15)

On the same fashion, if we define pBB the spot price after two bad states in a row, we have

pB ≥ δ (1 − π) p0BB (16)

Where the inequality comes from the fact that perhaps the generators will strictly prefer to sell in
this period.

Proof of Proposition 5.
In order to prove the proposition, we first need to prove the following lemma.

0
Lemma 6 We cannot have qH = D − R̄ in the current period and qH < D − R̄ in the subsequent
period if the renewables’ state is the same in both periods and there is no rain in the second period.

31
Proof. Fix θ > 0 and R ≥ 0. We will focus on the case of renewable’s good state, the proof
for the bad state is analogous. Assume by way of contradiction that we have a optimal policy
0 0
qH := qH (X, θ, R) = D − uh R and qH := qH (X − qH , θ, R), with qH > qH . Consider that the
ε
planner engage in a deviation of this optimal policy only on two periods, with qH = qH − ε and
ε 0 0
qH = qH + ε. If 0 < ε < D − uh R this policy is feasible. Denote by Cd the associated cost value.
We show that exists ε small enough such that Cε < C. Note that:

Cε − C = C(ε) − C(0) + δ(1 − π)((1 − γ)(CB (X − qH + ε) − CB (X − qH ))+

0 0
γ(C(D − uh R − qH − ε) − C(D − uh R − qH )))

Where in the last part of right hand side we used the fact that in our proposed deviation, nothing
changes from the third period onwards (that is, the state variables will be the same in the proposed
policy and in the deviation, regardless of the state of the world). Moreover, since C is decreasing
in the first argument we have:

0 0
Cε − C ≤ C(ε) − C(0) + δ(1 − π)γ(C(D − uh R − qH − ε) − C(D − uh R − qH ))

Since we need to show that Cε − C ≤ 0, it suffices to show that ∃ ε > 0, such that:
0 0
C(ε) − C(0) δ(1 − π)γ(C(D − uh R − qH ) − C(D − uh R − qH − ε))
<
ε ε

Both sides are positive since ∀x > 0 C 0 (x) > 0. The existence of such an ε comes directly from
0
the fact that D − uh R − qH > 0 and C 0 (D − uh R − qH 0
) > C 0 (0) = 0. Too see why, denote by
0 0
δ(1−π)γ(C(D−uh R−qH )−C(D−uh R−qH −ε))
a := C(ε)−C(0)
ε
, b := ε
and l = δ(1 − π)C 0 (D − uh R − qH
0
). By
these previous properties we have that ∀γ > 0, ∃δ > 0, such that:

ε < δ ⇒ (a < γ ∧ l − γ < b < l + γ) ⇒ b − a > l − 2γ

By picking γ = 4l and 0 < ε < δ we get the desired result. We get then that Cε < C which is a
contradiction with the definition of C.
Now assume by way of contradiction that on an arbitrary period t the planner chooses qt (Xt , θ, R) =
D − R̄. Fix s ∈ N, such that Xt < s(D − R̄). Consider a history of s periods of bad states for the
hydro power source (and the same states for the renewable source). Then by the previous lemma
we have by induction that qt+s = D − R̄, which is a contradiction, since it is necessarily true that
qt+s > Xt+s .

Proof of Theorem 1.
First we prove that the value function satisfies certain properties.

Lemma 7 For i ∈ {G, B}, V i (X, θ, R) is bounded, continuous, increasing and concave.

32
 Proof of Lemma 7. Let D be the space of bounded, continuous, non-decreasing and concave
functions defined on [0, θ] × [0, θ̄] × [0, D], endowed with the sup-norm. D is a complete space,
since it is a closed subset of the space of bounded and continuous functions. Since the Cartesian
product preserves completeness we have that D × D is also a complete space. Define the operator
T on D × D, T : (f, g) → (T G f, T B g):

T G f (X, θ, R) = max
0
{−C(D − uh R − X + X 0 ) + δ(πγf (θ, θ, R) + (1 − π)γf (X 0 , θ, R)+
X

π(1 − γ)g(θ, θ, R) + (1 − π)(1 − γ)g(X 0 , θ, R)}

s.t
max{0, X + uh R − D} ≤ X 0 ≤ X

and

T B g(X, θ, R) = max
0
{−C(D − ul R − X + X 0 ) + δ(πγf (θ, θ, R) + (1 − π)γf (X 0 , θ, R)+
X

π(1 − γ)g(θ, θ, R)) + (1 − π)(1 − γ)g(X 0 , θ, R)}

s.t
max{0, X + ul R − D} ≤ X 0 ≤ X

Denote by F G (X, θ, R, X 0 ) and F B (X, θ, R, X 0 ) the objective function from the problems above ,
respectively. Also define f¯(·) := πγf (·) + (1 − π)γf (·) + π(1 − γ)g(·) + (1 − π)(1 − γ)g(·). Define
the following correspondences:

ΓG (X, θ, R) = {X 0 ∈ R+ : max{0, X + uh R − D} ≤ X 0 ≤ X}
ΓB (X, θ, R) = {X 0 ∈ R+ : max{0, X + ul R − D} ≤ X 0 ≤ X}

Note that both correspondences are non-empty, compact-valued and convex-valued. Since in both
cases the functions on each side of inequality are continuous we conclude that both correspondences
are continuous.
(i) (T G f, T B g) is bounded and continuous: First note that since (f, g) ∈ D × D and
by assumption −C(·) is continuous and bounded then F i (X, θ, R, X 0 ) is linear combination of
continuous and bounded functions, so it is also continuous and bounded for i = {G, B}.Then
maximizing F i (X, θ, R, X 0 ) over the compact set Γi (X, θ, R) has a solution and by consequence
(T G f, T B g) is bounded. Moreover since Γi (X, θ, R) is compact valued and continuous we can apply
Berge’s Maximum Theorem to conclude that (T G f, T B g) is continuous.

33
 (ii) T G f is increasing: Fix (X̄, θ̄, R̄) ≥ (X, θ, R) with one >. Pick

x0 ∈ arg max F G (X, R, θ, X 0 )

X 0 ∈ΓG (X,θ,R)

Case 1: x0 ∈ ΓG (X̄, θ̄, R̄):

T g f (X̄, θ̄, R̄) = max −C(D − uh R̄ − X̄ + X 0 ) + δ f¯(X 0 , θ̄, R̄) ≥ −C(D − uh R̄ − X̄ + x0 ) + δ f¯(x0 , θ̄, R̄)
≥ −C(D − uh R − X + x0 ) + δ f¯(x0 , θ, R) = T g f (X, θ, R)

Where the last inequality inequality comes from the fact that −C(·) is strictly decreasing and f¯(·)
is non decreasing. Note that we will have a strictly inequality if X̄ > X or R̄ > R.
Case 2: x0 6∈ ΓG (X̄, θ̄, R̄): Then it must be the case that X̄ + uh R̄ − D > x0 ≥ 0:

T g f (X̄, θ̄, R̄) = max −C(D − uh R̄ − X̄ + X 0 ) + δ f¯(X 0 , θ̄, R̄) ≥ −C(0) + δ f¯(X̄ + R̄ − D, θ̄, R̄)
≥ −C(D − uh R − X + x0 ) + δ f¯(x0 , θ, R) = T g f (X, θ, R)

Where the last inequality comes from the fact that −C(0) is the maximum of −C(·).
(ii) T G f is Concave: Fix (X 0 , θ0 , R0 ), (X 00 , θ00 , R00 ), fix α ∈ [0, 1] and:

x0 ∈ arg max F G (X 0 , θ0 , R0 , X)
X∈ΓG (X 0 ,θ0 ,R0 )

x00 ∈ arg max F G (X 00 , θ00 , R00 , X)

X∈ΓG (X 00 ,θ00 ,R00 )

Define x000 = αx0 + (1 − α)x00 and (X 000 , θ000 , R000 ) = α(X 0 , θ0 , R0 ) + (1 − α)(X 00 , θ00 , R00 ). Then by
the constraints we have:

x000 ≥ max{0, α(X 0 +uh R0 −D)}+max{0, (1−α)(X 00 +uh R00 −D)} ≥ max{0, α(X 0 +uh R0 )+(1−α)(X 00 +uh R00 )−D

And we conclude x000 ∈ ΓG (X 000 , θ000 , R000 ). Then

max −C(D − uh R000 − X 000 + X) + δ f¯(X 000 , θ000 R000 ) ≥ −C(D − uh R000 − X 000 + x000 ) + δ f¯(x000 , θ000 , R000 )
≥ α(−C(D − uh R0 − X 0 + x0 ) + δ f¯(x0 , θ0 , R0 )) + (1 − α)(−C(D − uh R00 − X 00 + x00 ) + δ f¯(x00 , θ00 , R00 ))

Where the last inequality comes from the fact that −C(·) is concave and f¯(·) is concave. So we
conclude that:

T g f (X 000 , θ000 , R000 ) ≥ αT g f (X 0 , θ0 , R0 ) + (1 − α)T g f (X 00 , θ00 , R00 )

(iii) T B g is increasing and Concave: The proof is analogous to the case for T G f .
Using the previous results we can conclude that T:D × D → D × D. Since δ < 1 we can apply

34
the Blackwell’s Sufficient Conditions to conclude that T is a contraction in a complete space, then
we can apply the Banach fixed-point theorem to prove the existence of a unique fixed point of the
operator, (VG , VB ) ∈ D × D.
We can derive further properties if we combine the previous lemma with proposition 5. We
know that is always true that qH < D−R̄, then, by definition, it is always true that X 0 > X +R−D.

Proposition 7 For i ∈ {G, B}, V i (X, θ, R) is bounded, continuous, strictly increasing and strictly
concave.

Proof. By lemma 7 we know that V i (X, θ, R) is bounded, continuous, increasing and concave.
Then it suffices to show that the function is strictly increasing and strictly concave.
(i) V g is strictly increasing: The proof that V g is strictly increasing in X and R its analogous
to the proof of lemma 7, but applying Proposition 5 to get the strict inequality. The result that
V g is increasing in θ comes directly from the fact that from a given X, R and θ̄ > θ, ΓG (X, θ, R) ⊆
ΓG (X, θ̄, R) and the fact that V g is strictly increasing in X.
(ii) V g is strictly concave: The procedure is analogous. We first use proposition 5 to show
that for a fixed θ,V g is strictly concave in X and R, and use this fact to show that V g is strictly
concave in (X, θ, R).
The following two results will help us describe fully the social planner’s optimal policy. The
following lemma shows the optimal interior solution to the problem. We already know from
Proposition 5 that the planner always uses some thermal source to produce. The next result tells
us that off corners, the social planner’s optimal policy is to produce with the hydro plants such
that it equalizes the current thermal plant’s marginal cost with next period’s expected discounted
marginal cost.
∂V i (X,R)
Proposition 8 Fix i ∈ {G, B}, off corners we have ∂X
= C 0 (D − uh R − X + x0 ), where x0
is the optimal choice given the state (X, R).

Proof. We also have by assumption that C(·) is continuously differentiable on R+ then it will be
continuously differentiable on int ΓG (X, θ, R), with these and the previous results we can apply
the Benveniste-Scheinkman Theorem to get the result.
Now, consider the situation where there is not much water left on the reservoirs. In this case,
when there is no production of intermittent sources, the planner is better off producing using all
water left. This is true even though it will imply that in the subsequent period, if there is no rain,
all energy production will necessarily come from thermal sources and intermittent sources, at a
high cost. The following lemma describes optimal production in this case.

Lemma 8 There is a storage level X e > 0, such that in a state in which there is no production
of intermittent sources, for any amount of storage X in which X ≤ X, e all the water is used to

35
produce energy. Moreover, if it is also true that

C 0 (D − R) ≥ δ (1 − π) (γC 0 (D − R) + (1 − γ)C 0 (D)),

then there is also a threshold level Ye such that even in a state in which there is full production
of intermittent sources, for any amount of storage X in which X ≤ Ye , all the water is used to
produce energy.

Proof. Since C 0 (D) is the highest possible marginal cost, there is a storage level X
e > 0 such
that
C 0 (D − X)
e = δ (1 − π) (γC 0 (D − R) + (1 − γ)C 0 (D))

For any level X such that X < X,

e the following holds:

C 0 (D − X) > δ (1 − π) (γC 0 (D − R) + (1 − γ)C 0 (D))

Using the convexity of the cost function we get the result. The argument for the case of a good
state of intermittent is analogous.
We have thus far fully described the social optimum. Lemma 8 shows that there is a critical
amount of stored water for each state of intermittent production, such that the planner deploys all
reservoir in that period. Above these thresholds, we know that production the solution is interior
and given by Proposition 8.

9 Appendix B: Data Sources

Table 6: Data Sources: Brazil

Brazil 2018
Data Source url
Wind Capacity Factor Operador Nacional do Sistema Eletrico https://bit.ly/2BVRSu7
Composition of energy production Operador Nacional do Sistema Eletrico https://bit.ly/2BVRSu7
Hydro and Wind capacity Agencia Nacional de Energia Eletrica https://bit.ly/28INSwk

36
 Table 7: Data Sources: Uruguay
Uruguay 2017
Data Source url
Wind Capacity Factor Administracion Nacional de Usinas y Trasmisiones Electricas https://bit.ly/2XuLvqU
Composition of energy production Administracion Nacional de Usinas y Trasmisiones Electricas https://bit.ly/2EDZoO0
Hydro and Wind capacity Ministerio de Industria, Energia y Mineria https://bit.ly/2BZQfvL

Table 8: Data Sources: Argentina

Argentina 2017
Data Source url
Wind Capacity Factor Comision Nacional de Energia Atomica https://bit.ly/2Tql7Qu
Composition of energy production Comision Nacional de Energia Atomica https://bit.ly/2Tql7Qu
Hydro and Wind capacity Comision Nacional de Energia Atomica https://bit.ly/2Tql7Qu

References
[1] Stefan Ambec and Claude Crampes. Decarbonizing electricity generation with intermittent
sources of energy. Journal of the Association of Environmental and Resource Economists, 6,
2016.

[2] Stefan Ambec and Joseph A Doucet. Decentralizing hydro power production. Canadian
Journal of Economics/Revue canadienne d’économique, 36(3):587–607, 2003.

[3] Stefan Ambec and Yuting Yang. Climate policy and electricity trade. 2020.

[4] European Commission. Energy, Feb 2017.

[5] Claude Crampes and Michel Moreaux. Water resource and power generation. International
Journal of Industrial Organization, 19(6):975–997, 2001.

[6] Peter Cramton and Steven Stoft. A capacity market that makes sense. Electricity Journal,
(18):43–54, Aug 2005.

[7] Matthes F. Newbery D. Fabra, N. and A. Rudinger. The energy transition in europe: Lessons
from germany, the uk and france. working paper, 2020.

[8] Natalia Fabra and Gerard Llobet. Auctions with unknown capacities: Understanding compe-
tition among renewables. working paper, pages 812–830, 2020.

[9] Alfredo Garcia, James D Reitzes, and Ennio Stacchetti. Strategic pricing when electricity is
storable. Journal of Regulatory Economics, 20(3):223–247, 2001.

37
[10] Richard J Green and David M Newbery. Competition in the british electricity spot market.
Journal of Political Economy, pages 929–953, 1992.

[11] Bård Harstad. Pledge-and-review bargaining: From kyoto to paris. working paper, 2020.

[12] Carsten Helm and Mathias Mier. On the efficient market diffusion of intermittent renewable
energies. Energy Economics, 80:812–830, 2019.

[13] Paul L Joskow and Richard Schmalensee. Markets for power: An analysis of electrical utility
deregulation. MIT Press Books, 1, 1988.

[14] Paul L. Joskow and Jean Tirole. Reliability and competitive electricity markets. The RAND
Journal of Economics, 38, 2007.

[15] David Newbery. Missing money and missing markets: Reliability, capacity auctions and
interconnectors. Energy Policy, 94, 2016.

[16] Henry Thille Talat Genc and Khaled ElMawazini. Dynamic competition in electricity markets
under uncertainty. Energy Economics, forthcoming, 2020.

[17] Robert Wilson. Architecture of power markets. Econometrica, pages 1299–1340, 2002.

38