Learning by Doing on Both the Demand and the Supply Sides ISET

Document Sample
Learning by Doing on Both the Demand and the Supply Sides ISET Powered By Docstoc
					     Learning-by-Doing on Both the Demand and the Supply Sides:
    Implications for Electric Utility Investments in a Heuristic Model

                               John A. "Skip" Laitner
                        EPA Office of Atmospheric Programs
          1200 Pennsylvania Avenue NW, MS 6201-J, Washington, DC 20460
                                 o: (202) 564-9833
                                 f: (202) 565-2147

                                    Alan H. Sanstad
                         Lawrence Berkeley National Laboratory
                            #1 Cyclotron Road, MS 90-4000
                                 Berkeley, CA 94720
                                  o: (510) 486-6433
                                   f: (510) 486-6996


A growing literature is focusing on the phenomenon of “learning-by-doing” in the
context of energy supply technologies: As new technologies enter the marketplace, and
as experience is gained in both their production and use, costs tend to decline with each
successive doubling of investment or production. This work builds both on the long-
established literature on learning “curves” in production in a variety of industries and on
the classic work by Arrow in a general equilibrium context. A number of studies have
indicated that the effects of including learning-by-doing in energy forecasting and
simulation modeling may be substantial relative to modeling with only “autonomous” or
“exogenous” technical change. However, much less attention has been paid to the
implications of learning-by-doing for demand-side technologies. Given that these are
durable goods and thus subject to learning effects on production, there will be underlying
cost declines that affect the end-use cost of energy services, and omitting such effects
may introduce a bias into technology forecasts that incorporate learning-by-doing. This
paper explores the implications of this observation through the application of a heuristic
model that captures the anticipated electricity service demand within the United States
over the next 30 years. We examine how including demand as well as supply-side
learning could impact investment decisions within the U.S. electric utility industry.

Energy-economic simulation modeling and forecasting, of both general and partial

equilibrium varieties, has traditionally embodied technological change as “autonomous”

or “exogenous,” formally, as a function of time only. The best known example is the so-

called “AEEI” or “autonomous energy efficiency index,” a parameter that appears

generically in virtually all energy-related simulation models.        With this type of

representation, technological change is essentially “manna-from-heaven:” it is unaffected

both by changes in relative prices and by policy actions on the part of the government. It

is well-known that the magnitude of this type of parameter substantially affects estimated

costs of policies to reduce energy demand or carbon emissions. Thus, the “central

tendency” in energy-economic modeling has been, in effect, to formally put off limits to

policy one of the most important channels for reducing the costs of energy efficiency or

carbon abatement.

The significance of this limitation has become increasingly apparent as work has

expanded on energy policy analysis related to climate change. There has thus been a

substantial increase in recent years in research aimed at implementing alternative

representations of technological change in energy and climate policy modeling. This

work has focused on incorporating plausible mechanisms for technological change that

deviate from the standard autonomous paradigm, and on examining their policy

implications. A key such mechanism is that of “learning-by-doing (LBD):” the long-

observed phenomenon of declining costs of production of technologies with experience
on the part of manufacturers or as a function of cumulative investment or production.

LBD has been documented extensively over many decades in technology assessment

research beginning with the work of Wright (1936). (See, e.g., Argote and Epple 1990.)

In the general equilibrium setting, the classic work by Arrow (1962) has been one of the

starting points for contemporary research on endogenous economic growth, catalyzed by

the work of Romer (1986), which re-conceived and dramatically extended Arrow’s


Recent work on LBD specifically as it applies to energy technology has been carried out

using either technology-cost and optimal planning or general equilibrium approaches.

With few exceptions, however, this work has focused on energy supply technologies. It

is known however, that the production of energy-using consumer durables also manifests

learning effects.   As Watanabe et al. (2002) and others have noted, omitting this

‘demand-side’ effect may introduce a bias in LBD analyses of energy-related technology,

particularly when these are conducted in a general equilibrium setting.

This paper presents a simple analysis in which LBD is simultaneously represented on

both the supply and demand sides, in the context of a prototypical market for electricity.

Our results confirm that including demand-side learning effects indeed can substantially

affect cost estimates arising in LBD analyses. With the caveat that this is a preliminary

and exploratory analysis, this indicates the importance of extending the scope of LBD

and energy technology research to include demand-side detail
The paper is organized as follows. We begin by sketching recent research on LBD and

energy supply technology, and then give a brief summary of empirical findings on LBD

and consumer durables. Next we present our simple electricity model, with a discussion

of the structural and empirical assumptions including parameter values and calibration

approach. We then summarize the empirical results, distinguishing in turn the effects, on

cost estimates, of including supply-side LBD only and of including both supply and

demand side LBD. The paper concludes with remarks on policy implications.

Background: LBD and Energy-Related Technologies.

As noted above, recent research on LBD in the context of energy modeling and policy

analysis has included both direct technology-cost studies, estimating learning parameters

in the classic sense, and studies in which the implications of LBD are examined in an

optimal planning or general equilibrium setting. In the technology-cost category, Wene

(2000) presents estimates of learning rates for several technologies, including wind

turbines, photovoltaics, district heating, and natural gas combined cycle. The latter is

examined in detail by Colpier and Cornland (2002), while Molburg et al. (1995) analyze

LBD in a number of generation technologies.

General equilibrium and optimal planning studies of LBD and energy have focused both

on how costs of, for example, meeting GHG emissions targets are affected by LBD as

well as how the timing of abatement may be affected, whether in meeting given targets or
in an optimizing (cost-benefit) context, relative to benchmarks in which technological

change arises autonomously. Goulder and Mathai (2000) find that, while the effects of

LBD on the timing of abatement are ambiguous, it results in substantially lower costs of

meeting emissions targets, and in a cost-benefit analysis also increases the optimal level

of cumulative abatement. In a general equilibrium analysis, Manne and Richels (2002)

similarly find that, while LBD only weakly affects optimal abatement timing, it does

imply substantially lower costs of abatement. (See also Manne and Baretto 2002.)

Rasmussen (2001) presents a general equilibrium analysis in which renewable energy

technology is subject to LBD, and also finds a substantial cost-reducing effect.

There has been considerably less research on LBD and consumer durables than on other

technology categories, but enough to establish the presence of learning effects. Bass

(1980) presents econometric estimates for several consumer durables (air conditioners,

clothes dryers, dishwashers, refrigerators, and televisions), finding learning rates

(specifically, percent decreases in cost for a 1 percent increase in cumulative production)

between 0.1 and 0.4, while Boston Consulting Group (1972) reported learning rates for

electric and gas ranges and televisions between 0.1 and 0.5. In a very interesting recent

study, Newell (2000) estimated learning rates for consumer room and central air

conditioners and gas water heaters controlling for product quality, and found them to be

approximately 0.4.
Some specific notation will be useful to present further data on learning effects as well as

the model in the following section. The functional form of the learning curve can be

expressed as:

                                   Costt =Cost0*(1−a) d                              (Eq. 1)

where Costt is the cost of production at time t, Cost0 is the initial cost of production in the

base year zero, a is reduction in cost for each doubling of cumulative production, and

finally, d is the number of doublings in the cumulative output of a given commodity. The

expression 1-a can also be thought of as the ‘Progress Ratio.’ If a has a value of 10

percent, for example, then each doubling of cumulative output leads to a cost that is 90

percent of the previous value.

For purposes of our policy simulations, it is more convenient to express the learning

                                 Cost t = Cost 0 * Output b                          (Eq. 2)

curve as a power function:

where Output is an index of cumulative number of technology units produced over a

period divided by the cumulative units produced in the first year, and b is a learning

parameter that measures the rate those costs are reduced as cumulative output increases.

Thus the progress ratio can be written as approximately 2b. As an example, a total of

350.4 million electronic ballasts for florescent lighting were installed in the United States

                      Cost t = $36.47 * (350.4 / 0.4) −0.1909 = $10.09               (Eq. 3)
in the period 1986 through 2001. The number of ballasts installed by 1986 was 430,000

in 1986 with a cost of $36.47 per ballast (in 1996 dollars). Based upon previously

estimated data, we can apply Equation 2 for this technology to obtain Equation 3, where

the ‘learning parameter’ is -0.1909.1 These values are derived from data found in Laitner

(2002) that are in turn based upon a variety of other sources documenting the influence of

cumulative production upon technology costs. A mathematical manipulation shows that

the 2001 unit cost in the above equation is the result of an 87.6 percent progress ratio.

Additional examples of progress ratios for various technologies are shown in the Table 1

(Laitner, 2002), in which progress ratios range from 67 to 97 percent. A so-called

‘mature’ technology such as the magnetic ballast shows a 97 percent progress ratio,

indicating a slow rate of cost decline. By contrast, a more advanced technology for the

same end use, in this case the more efficient electronic ballast, demonstrates a 88 percent

progress ratio. The pollution control technologies in the table, including CFC substitutes

and scrubbers, appear to cluster around the 90 percent benchmark.

Table 1. Examples of Typical Progress Ratios for Selected Consumer Durable

                                                  Year 1          Cumulative          Cost     Progress
    Technology                    Period       Production          Production        Index      Ratio
    Ford Model T Auto           1909-1923         15,741            8,028,992        0.290       87%
    Integrated Circuits         1962-1968     4 million units   828 million units    0.047       67%
    CFC Substitutes             1988-1999      100,000 tons      3,871,000 tons      0.690       93%
    Scrubbers                   1987-1995        65.8 GW            84.3 GW          0.941       89%
    Photovoltaic Cells          1971-2000        0.1 MW            1451.4 MW         0.042       72%
    Magnetic Ballasts           1977-1993      29.4 million       629.3 million      0.897       97%
    Electronic Ballasts         1986-2001         431,000       350 million units    0.277       88%

 . The parameter -b can generally be found by the expression Ln(1-a) / Ln(2). For a more complete review
of the mathematics, see, Molburg, et al (1995). In this instance, Ln(1-0.1239) / Ln(2) = -0.1909.
 Refrigerators               1980-1998     5.1 million     126.3 million     0.556     88%
 Freezers                    1980-1998     1.8 million      26.1 million     0.374     78%
 Clothes Washers             1980-1998     4.4 million     104.7 million     0.536     87%
 Electric Clothes Dryer      1980-1998     2.5 million      61.0 million     0.557     88%
 Gas Clothes Dryer           1980-1998     0.7 million      18.2 million     0.593     90%
 Dishwasher                  1980-1998     2.7 million      69.7 million     0.450     84%
 Room Air Conditioner        1980-1998     2.4 million      63.3 million     0.478     85%
 Selective Window Coatings   1992-2000   4.8 million m2   157.4 million m2   0.394     83%

A Spreadsheet Model of Electricity Technologies and Choices

In order to explore the role of learning in electricity choices, we suppose that the United

States plans to meet total electricity demands generally following the reference case

trends of the AEO2002 (Energy Information Administration 2001).                In this case,

however, we extend the analysis beyond the year 2020 to the year 2032 to simulate a full

depreciation of existing capital stock by the year 2032. In the year 2002, total electricity

consumption is estimated to be 3,502 billion kWh. Assuming this will increase at a rate

of about 1.9 percent per year through 2032, total demand will increase to 6,104 billion


To meet these reference case demands, we suppose that there are five technologies to

meet both existing and new electricity demands.

   1. Existing capital stock. This is the average type of unit that is on line in 2002;

       although dominated by a mix of fossil fuel technologies, it also includes a mix of

       hydroelectric, renewable, and nuclear technologies. The assumption is that the

       existing capital stock depreciates by 3.3 percent annually beginning in 2003 such
         that the existing stock is completely out of service by 2032 and replaced by one of

         the four remaining technologies. It is not subject to learning.

   2. Defender Technology. This technology is the same as the existing capital stock,

         but it is subject to learning and competes to replace existing capital stock or to

         meet new electricity demand through 2032.

   3. Challenger Technology. The initial challenger technology which initially has

         higher costs, but is also assumed to benefit from a slightly higher rate of learning.

   4. Advanced Challenger. This is a more efficient technology, but subject both to

         higher initial costs and rates of learning.

   5. Demand-Side Efficiency Investments. This is category of end-use technologies

         that potentially will impact reference case demand for electricity. While the

         reference case demand-side technologies are assumed to cost less than the busbar

         cost of existing capital stock and defender technologies ― and hence, penetrating

         as part of the normal reduction in electric intensity of the economy, these

         additional demand-side technologies are assumed to have initial costs similar to

         the Challenger Technologies with learning rates comparable to the Advanced

         Challenger technologies.

Further details and cost characteristics of these technologies are provided in Table 2

below.     Also included are the characteristics of the transmission, distribution, and
administrative (TD&A) costs associated with the delivery of a kWh to end users.

Together with the busbar costs of electricity, this allows us to estimate the total annual

electricity bill for any given year.

Table 2. Illustrative Values of Cost Parameters

                               1             2                3           4         5         6
                            Existing      Defender        Challenger   Advanced   Demand   T,D&A
Static cost coefficient
                             0.040            0.025         0.025       0.025     0.025    0.005
Initial learning cost
                             0.000            0.015         0.035       0.075     0.035    0.021
coefficient ($/kWh)
Initial accumulative
                               1               1              1           1         1        1
experience (billion kWh)
Progress ratio                0.00            0.95           0.90        0.85      0.85     0.95

In this analysis, the year 2003 average cost of electricity is assumed to be $0.0664 per

kWh (in 2000 dollars). If existing capital stock is 0.040/kWh as shown in Table 2, this

implies that transmission, distribution, and administrative costs are $0.0264.                The

nation’s total electricity bill is $241 billion (also in 2000 dollars). New capital stock is

introduced using the Market Share algorithm as a function of annualized costs per kWh:

                                               COSTkt v
                                     MSkt =    J                                           (Eq. 4)
                                              ∑ COST
                                              j =1


MSkt = market share of technology k at time t
COSTkt = amortized capital and operating costs of technology k at time t

v = variance parameter representing cost homogeneity

j = technologies competing to provide the same service as k.

The function MSkt is a logistic curve whose slope is determined by a variance parameter,

v. A high value for v, such as 100, means that technology with the lowest cost captures

almost all of the new equipment stocks, as would occur with a linear programming

model. An extremely low value, such as 1, means that new equipment market share are

distributed almost evening among all competing technologies, even if their annual costs

differ significantly. An extremely high value, such as 10, means that the most cost

effective equipment gains a proportionately higher market share.           For example, a

technology with a 25 percent cost advantage would grab 90 percent of market share. In

this exercise, we adopt a value of 4. In this case, a technology with a 25 percent cost

advantage would grab 71 percent of the market share.


Figures 1, 2, and 3 go about here


Numerical Results

Assuming no learning and no efficiency in the reference case (see Figure 1), by 2032 the

defender technology provides an estimated 82 percent of total electricity demand while
the challenger and advanced technologies provide 16 and 2 percent, respectively. The

total electricity bill is estimated at $430 billion in that same year. With the previously

described learning assumptions incorporated into the supply-side mix, the defender share

in 2032 declines to 56 percent (see Figure 2). The challenger and advanced technology

shares increase to provide 25 and 19 percent, respectively. The total electricity bill, in

this case, declines to $390 billion in the year 2032.

However, if LBD is assumed to apply to the demand-side as well as the supply-side (see

Figure 3), then the defender market share further declines to 39 percent. The challenger

and advanced technology shares also decline to 17 and 12 percent, respectively. Under

these assumptions, efficiency investments capture 32 percent of the market share. In

effect, the demand-side efficiency investments reduce electricity consumption by 32

percent compared to the reference case. The total expenditures for energy services,

including the amortized costs of both supply and demand-side investments, decline to

$286 billion.

While these results are intuitively satisfying, we also examined the impact of different

assumptions with respect to initial costs, rates of learning, and the variance parameter in

the market share algorithm. Although the specific outcomes obviously differed in each

of the sensitivity run, in each case the inclusion of end-use energy efficiency technologies

significantly affected the both the market share and the costs associated with each of the

scenarios. Hence, the conclusion remains that any given scenario must include the
effects of LBD on both the demand and the supply sides of the equation to avoid biased


Discussion and Policy Implications

The calculations presented above apply the standard simple formula relating cost directly

to cumulative production. It is important to note, however, that this is a reduced form for

a much richer and more complicated set of phenomena that can result in declining unit

cost for a given technology (i.e., cost per unit of useful energy service delivered). These

include: (a) new knowledge that continuously flows into the processes by which the

technology is produced; (b) economies of both scale and scope that can be achieved with

increasing levels of production of the technology; (c) ‘learning by doing’ in the

manufacture of the technology simultaneously with ‘learning by using’ the technology,

and the feedback between these, even with no change in the physical capital stock; (d)

improvements induced in ancillary technologies that render system integration more

efficient; (e) improvements in the knowledge, skills and productivity of associated

engineering and labor; and, (f) the proliferation of service and distribution networks that

reduce the cost to consumers using the new technologies. 2

Including LBD captures, at least in a simple way, some of this richness in technological

change and diffusion. As our prototypical model and several studies cited above show,
this re-configuration alone in simulation models consistently results in lower cost

estimates than are obtained with the standard representation of technological change.

Parallel recent work on incorporating technological change due to research and

development in energy-economic models points in the same direction (e.g., Popp 2002).

Overall, this emerging body of work is demonstrating that generally accepted estimates

of the costs of large-scale carbon abatement require downward revision. Moreover, not

just direct cost estimates but the identification of the fundamental economic

characteristics of the policy problem change when LBD and related phenomenon are

recognized. The standard paradigm for computable general equilibrium models applied

to energy economics and carbon policy is that the competitive equilibrium is Pareto

optimal. With the inclusion of LBD specifically or endogenous technological change

more generally, however, this benchmark is inappropriate: these mechanisms imply an

underlying market failure even in the absence of environmental externalities (formally,

due to non-convexities that arise when these alternative representations of technological

change are introduced). This points toward a need for further model development, both

theoretical and empirical, related to technological change, and the absorption of the

results into the policy process.

More prosaically, future estimates of energy and pollution control technology costs, even

short of full general equilibrium simulation, should anticipate some decline in the cost of

these technologies over time relative to what would be projected with only autonomous

technological change. Policies that increase market share would accelerate this process

  To give a simple example of possible synergies, while the cost per unit of capacity of a wind
electricity generator can be reduced, technological improvements could simultaneously increase the
and serve to overcome the underlying market failure due to learning effects. Given the

prima facie short run advantages that may be enjoyed by less efficient technologies, i.e.,

when viewed strategically by investors, the importance of policy induced technological

progress can be highlighted in dynamic modeling analysis that represents learning and

shows the long run economic advantages of the advanced technologies.


Mr. Sanstad’s participation in this modeling exercise was supported by the Office of

Atmospheric Programs of the U. S. Environmental Protection Agency and prepared for

the U. S. Department of Energy under Contract No. DE-AC03-76SF00098. The opinions

expressed in this article do not necessarily reflect those of the U.S. Environmental

Protection Agency or the U.S. Government. All errors and misuse of statistics remain the

sole responsibility of the authors.


Argote, Linda, and Dennis Epple. 1990. “Learning Curves in Manufacturing.” Science

247 (4945), February 23, 920-924.

capacity factor, thereby further reducing the unit costs of electricity that it generates.
Arrow, Kenneth J. 1962. “The Economic Implications of Learning by Doing.” Review of

Economic Studies 29, 155-173.

Colpier, Ulrika Claeson, and Deborah Cornland. 2002. “The economics of the combined

cycle gas turbine – an experience curve analysis.” Energy Policy 30, 309-316.

Energy Information Administration, 2001. Annual Energy Outlook with Projections to

2020, DOE/EIA-0383(2002), U.S. Department of Energy, Washington, DC, December.

Goulder, Lawrence H., and Koshi Mathai. 2000. “Optimal CO2 Abatement in the

Presence of Induced Technological Change. Journal of Environmental Economics and

Management 30, 1-38.

Laitner, John A. “Skip”. 2002. Unpublished Data on Experience Curves. U.S.

Environmental Protection Agency, Washington, DC, October.

Manne, Alan S., and Leonardo Barreto. 2002. “Learning-by-Doing and Carbon Dioxide

Emissions.” Working paper, Stanford University, Palo Alto, CA.

Manne, Alan S., and Richard G. Richels. 2002. “The Impact of Learning-by-Doing on the

Timing and Costs of CO2 Abatement.” Working paper 02-8, American Enterprise

Institution/Brookings Institution Joint Center for Regulatory Studies, May.
Molburg, John C., and Gale A. Boyd, J.D. Cavallo. 1995. “A Learning Functions of New

Electric Power Generating Technologies.” Working paper, Argonne National Laboratory,

Argonne, IL.

Newell, Richard G. 2000. “Incorporation of Technological Learning into NEMS Building

Modules,” Draft report to the Energy Information Administration, September.

Newell, Richard G., and Adam B. Jaffe, Robert N. Stavins. 1999. “The Induced

Innovation Hypothesis and Energy-Saving Technological Change.” Quarterly Journal of

Economics 114(3), 941-975, August.

Popp, David. 2002. “ENTICE: Endogenous Technological Change in the DICE Model of

Global Warming.” Working paper, Syracuse University, April.

Rasmussen, Tobias. 2001. “CO2 abatement policy with learning-by-doing in renewable

energy.” Resource and Energy Economics 23, 297-325.

Romer, Paul. 1986. “Increasing Returns and Long-run Growth.” Journal of Political

Economy 94 (5), 1002-1037.

Wene, Clas-Otto. 2000. Experience Curves for Energy Technology Policy. International

Energy Agency, Organization for Economic Cooperation and Development, Paris.
Watanabe, Chihiro, et al. 2002. “Diffusion Trajectory of Self-Propagating Innovations

through Interactions with Institutions.” Working paper, Tokyo Institute of Technology,

Tokyo, Japan.

Wright, Theodore P. 1936. “Factors Affecting the Cost of Airplanes.” Journal of

Aeronautical Sciences 3 (4), 122-128.
                        Figure 1.

                                    Exploring the Role of Learning on Technology Penetration - Reference Case



Electricity Use (TWh)


                        3,000                                                         Electricity Expenditures in 2032: $430 Billion



                            2002         2005            2008       2011       2014     2017       2020        2023     2026       2029        2032

                                Existing Capital Stock          Defender Technology    Challenger Technology     Advanced Challenger      Efficiency
                          Figure 2.

                                      Exploring the Role of Learning on Technology Penetration - With Learning



Electricity Use (TWh)


                                                                                  Electricity Expenditures in 2032: $390 Billion



                           2002          2005        2008       2011       2014     2017       2020        2023     2026       2029        2032

                                Existing Capital Stock      Defender Technology    Challenger Technology     Advanced Challenger      Efficiency
                           Figure 3.

                                     Exploring the Role of Learning on Technology Penetration - With Learning
                                                     and Demand-Side Efficiency Investments


Electricity Use (TWh)



                                                                                  Electricity Expenditures in 2032: $286 Billion


                           2002          2005        2008       2011       2014     2017       2020        2023      2026       2029       2032

                                Existing Capital Stock      Defender Technology    Challenger Technology      Advanced Challenger      Efficiency

Shared By: