Implications of uncertainty in technology cost projections for least-cost decarbonized electricity systems

Summary Plans for decarbonized electricity systems rely on projections of highly uncertain future technology costs. We use a stylized model to investigate the influence of future cost uncertainty, as represented by different projections in the National Renewable Energy Laboratory 2021 Annual Technology Baseline dataset, on technology mixes comprising least-cost decarbonized electricity systems. Our analysis shows that given the level of future cost uncertainty as represented by these projections, it is not possible to predict with confidence which technologies will play a dominant role in future least-cost carbon emission–free energy systems. Successful efforts to reduce costs of individual technologies may or may not lead to system cost reductions and widespread deployments, depending on the success of cost-reduction efforts for competing and complementary technologies. These results suggest a portfolio approach to reducing technology costs. Reliance on uncertain cost breakthroughs risks costly outcomes. Iterative decision-making with learning can help mitigate these risks.


Uncertainty of money spending on different technologies considering uncertainties in future cost projections under full emission reduction constraint
Cost ($ per kWh of mean demand)

INTRODUCTION
Stabilizing global mean temperature and reducing adverse consequences of the climate change caused by increasing concentration of atmospheric carbon dioxide (CO 2 ) motivate rapid decarbonization of the entire economy. 1,21][12][13][14] To economically decarbonize the electricity sector, no-or low-carbon emission generation technologies with reduced costs are needed to avoid emissions from fossil fuel sources (e.g., coal, oil, and natural gas).Studies of decarbonized electricity systems often consider a portfolio of low-carbon emission technologies with prominently different characteristics, 4,9,[15][16][17][18][19][20][21][22][23][24] and often use an ad hoc set of assumptions for future costs as chosen by the author teams.Among these technologies, variable renewables-mainly onshore wind (denoted as wind to distinguish from the offshore wind hereafter) and solar photovoltaics (solar) generation-harness renewable energy inputs from nature and have no fuel cost.0][31][32][33] In contrast, the most notable low-carbon emission firm technology, the nuclear power plant, is less strongly affected by weather conditions, and can generate stable power outputs.However, many other concerns (safety issue, waste disposal, etc.) have resulted in increased costs and moderated policy support for nuclear power, and led to early retirements of nuclear power plants in some regions. 34,35odeling studies that explore pathways to cost effectively decarbonize electricity systems rely on cost estimates or learning-rate assumptions for technologies as inputs to models.7][38] Whereas nearcurrent technology cost estimates are similar to one another, future cost projections diverge and are much more uncertain.Projected technology costs made by different groups or different trajectories within the same group vary due to differences in assumptions, such as levels of confidence in future technology innovations.Projected technology costs made by the same group also vary with time as a result of updated underlying assumptions with improved understanding.][47][48][49] Other studies considering cost reduction possibilities of more technologies often show additional benefits by including low-carbon emission firm generation (e.g., nuclear, biopower, and geothermal). 18,20,21][52][53][54] For example, Neumann and Brown (2023) considered cost uncertainties on renewable technologies and identified ranges of cost-efficient capacity expansion plans for European electricity systems; Pilpola and Lund (2019) 40 applied a Monte Carlo approach on Finnish energy systems and emphasized the importance of addressing input uncertainties in future low-carbon-emission energy system planning.Most of these studies, however, embedded the cost uncertainty within a wide range of sources of uncertainty, such as the level of electricity consumption and renewable resource availability, and they focused primarily on system-level behaviors including the annual system cost, CO 2 emissions, and total power supplied.It remains unclear about the sole impact of future cost uncertainty on system dynamics and the resulted technologymix in least-cost carbon-emission-free energy systems.Given the urgent need to eliminate carbon emissions and the fact that power sector capital equipment, once built, will remain in place for a long time and have long-lasting effect, it is important to understand the impact of uncertainty in future technology cost projections on the composition of least-cost decarbonized electricity systems.
In this study, we investigate how different technology cost projections could affect simulations of the technology mix of future decarbonized electricity systems.As we will show, given the level of future cost uncertainty, it is not possible to predict with confidence which technologies will play a dominant role in a future least-cost carbon emission-free energy system, and cost reduction in certain technologies do not necessarily guarantee the increased deployments of those technologies.
Here, we use the macro energy model (MEM). 23,24,47,49,55MEM is constructed as a linear-optimization model that considers only technoeconomic factors.The objective function for minimization is the total system cost associated with capacity expansion and power dispatch of a portfolio of generation and storage technologies (see Figure 1 for technologies considered in the main cases).A portfolio approach is thought to provide economic and technical benefits 56,57 and here to better capture the impact of cost variation of a wider range of technologies.Electricity provided to the exchange node is used to meet electricity demand, or curtailed as appropriate.Here, we conduct a series of single-year optimizations with hourly time resolution in a ''greenfield'' setting (i.e., no assumption of pre-existing capacity) and require that 100% electricity demand be satisfied at each hour.To facilitate understanding, we normalize the demand profile by dividing annual mean demand on each hourly step (i.e., annual mean demand is 1 kWh after normalization).Hourly demand and wind and solar generation potentials are calculated for the US (see STAR Methods).A fixed cost is associated with all technologies to represent the fixed capital investment (Table S1).This includes the purchase and installation costs and fixed operation and maintenance (O&M) costs.A variable cost is specified for gas, gas with carbon capture and storage (gas-with-CCS), nuclear, and biopower.This variable cost includes variable O&M costs and fuel costs as appropriate.For nuclear, we assume that the nuclear reactor must be operated at constant rates, and thus operation and fuel costs (i.e., variable costs) are added to the fixed cost as dollar per unit capacity.
A common problem in assessing effects of cost uncertainty on future least-cost system is attaining some consensus on the probability distribution of future technology costs or learning rates.Here, we aim to illustrate the role of cost uncertainty without making claims regarding future probabilities ourselves.Therefore, we adopt near-current (i.e., year 2019) and future (e.g., year 2050) cost estimates for technologies in the main cases based on the National Renewable Energy Laboratory (NREL) 2021 Annual Technology Baseline (ATB) dataset. 37We consider single cost estimates for all technologies for the year 2019 cases, and three cost estimates for future-year cases for each technology except for gas, biopower, and nuclear, which have one cost projection in the NREL 2021 ATB dataset.With these reduction pathways, we aim to represent the advanced, moderate, and conservative technology innovation pathways, which lead to low, middle, and high future cost levels, respectively.Numbers from the 2021 ATB dataset are converted to cost inputs for MEM in a consistent manner for all technologies.That is, we do not generate any cost estimates ourselves, but rely entirely on cost projections from the NREL ATB dataset.To isolate the impact of different cost projections, we compare simulations using the same electricity demand and resource profiles.We analyze six technologies with three cost projections (i.e., gas-with-CCS, solar, wind, offshore wind, geothermal, and battery storage).This leads to a total of 729 possible cost combinations (243 combinations under 100% emission reduction scenarios since we focuse on eliminating carbon emissions from gas and gas-with-CCS).The wide range of cost combinations facilitates our understandings of the system dynamics under different technology innovation scenarios.Where a probability assessment is required for illustrative purposes, we consider these cost projections to be equally probable.Caveats of these assumptions are discussed in limitations of the study.
Our analysis provides an idealized and transparent framework.We do not claim to have predictive skill in projecting technology costs.To illustrate the influence of cost uncertainty, we rely on the NREL ATB cost projections, and focus on the fundamental understanding of systematic dynamics.Many factors could affect future electricity costs and system planning, such as the availability of innovative low-carbon emission technologies, changes in global and domestic policy conditions, market responses, etc., which are beyond the scope of this study.Numerical values presented here might change easily, while qualitative conclusions about the influence of cost uncertainty on future least-cost systems are likely robust.We provide transparent analysis with full details disclosed, and hope to inspire discussions on such topics and studies that apply more sophisticated assumptions and advanced approaches.
In the Results section, we focus on results from our main cases that consider technologies listed in Figure 1 and technology cost estimates from the NREL 2021 ATB report.In the limitations of the study section, we further cover issues such as lower future cost estimates made for renewable technologies in the past, 25,58,59 impact of different discount rates, impact of including long-duration storage and direct air capture, using more recent future technology cost projections, other demand profiles, and other caveats associated with our stylized modeling framework.Our main findings are summarized in the discussion section.

Electricity system under representative cost projections
We first discuss results under the NREL ATB year-2019 cost levels for the US, where we gradually enforce carbon emission reduction constraints that remove fossil fuel sources (i.e., gas and gas-with-CCS).More information can be found in Text S1.In our simulations, gas and gas-with-CCS are the primary sources of electricity dispatch when there is small cost on carbon emissions (i.e., low emission reduction constraints) (Figures S1 and S2), because they are the lowest-cost way to meet electricity demand given the 2019 cost levels.As emission reduction constraints tighten, more electricity comes from wind and solar, with flexible fossil fuel generations filling gaps between variable electricity supply and demand.Under very deep emission reduction constraints (e.g., >90% emission reduction constraints), nuclear and biopower can be least-cost providers of reliable electricity, with nuclear providing constant and stable electricity generation, and wind and solar, supported by an increasing amount of biopower and battery storage, address variability in demand.When biopower is removed from the system, nuclear becomes more competitive and dominates the deep decarbonization scenarios (Figure S3) due to the high variability of wind and solar, and high cost of batteries; when nuclear is removed as well, curtailments from wind (both onshore and offshore) and solar increase substantially under deep emission reduction constraints along with system costs.These results are consistent with previous simulations using the same model and near-current cost estimates from the Energy Information Administration (EIA). 23Results using year 2016 to year 2018 demand, with wind and solar potentials are shown in Figure S4.
Projected year-2050 fixed costs of all technologies show substantially different cost reductions relative to the 2019 estimates.For example, ratios of fixed costs between 2050 and 2019 are >85% for technologies such as gas, biopower, nuclear, and geothermal under high future cost projection (i.e., conservative technology innovation) trajectories, suggesting modest cost reductions.Meanwhile, these fixed cost ratios are <40% for technologies including wind, solar, storage, and geothermal under low future cost projection (i.e., advanced technology innovation) trajectories (Figure S5 and Table S1), indicating great innovations have been achieved for these technologies.
We analyze six technologies with three cost projections (i.e., gas-with-CCS, solar, wind, offshore wind, geothermal, and battery storage).This leads to a total of 729 possible cost combinations (243 combinations under 100% emission reduction scenarios because gas-with-CCS cannot participate).To facilitate analysis of different cost assumptions, four illustrative 2050 cost combination cases are selected and compared in Figure 2, in which: all these six technologies follow the high future cost projections to represent conservative innovations for all technologies (''HighCost''); only geothermal achieves the advanced innovations and reaches its low cost projection while other technologies remain at high costs projections (''HighCost_LowGeo''); both wind and geothermal reach their low cost projections (''HighCost_Low-WindGeo''); and all technologies achieve their advanced innovations and reach their low cost projections (''LowCost'').HighCost and  LowCost are chosen because they represent NREL ATB's most optimistic and pessimistic technology innovation scenarios; HighCost_Low-Geo is used to highlight the case where cost ratios between renewables and firm technologies are larger and geothermal dominates renewables under deep emission reduction constraints; and HighCost_LowWindGeo is used to represents the case where both renewables and geothermal contribute substantially under deep emission reduction constraints.
Several features are noticeable across these cost combination cases.First of all, all four cases share similar behaviors as they move from low to modest emission reduction constraints.For example, gas consistently acts as the least costly way to meet electricity demand when no emission reduction constraint is implemented.Under modest emission reduction constraints (e.g., <80% emission reduction constraint), low-carbon emission firm technologies cannot compete with wind and solar.Substantial cost reductions in wind and solar in 2050 lead to decreased use of gas-with-CCS compared to the 2019 results.Competition between wind and solar depends initially on their relative magnitudes of cost reductions.As emission reduction constraints are enhanced, wind becomes more important (Figures 2 and S6).
The four scenarios with representative cost combinations examined here show distinguishable outcomes under deep emission reduction constraints (e.g., >90% emission reduction constraint).For the HighCost case, because the 20% cost reduction for wind and >40% cost reduction for solar are larger than cost reductions for firm technologies (less than 15% for nuclear and geothermal; both of them also have substantially larger costs per kWh than wind and solar at the 2019 level), wind and solar are much more attractive and outcompete firm generations in deep emission reduction scenarios, with dispatches from biopower to fulfill flexibility purposes.In HighCost, nuclear is rarely used even under the 100% emission reduction constraint because costs of nuclear are projected to decline less than the costs of most competing technologies (details depend on the year of hourly demand and wind and solar profiles used for optimizations, see Figure S4).
When only geothermal achieves advanced cost innovations and reaches its low future cost projection (i.e., to 27% of current costs, High-Cost_LowGeo, Figure 2B), geothermal replaces much wind and solar, and dominates systems under deep emission reduction constraints.If wind also achieves its low future cost projection (i.e., to 32% of current costs, HighCost_LowWindGeo, Figure 2C), the market share for geothermal is markedly reduced.If all competing technologies achieve their low future cost projections (LowCost, Figure 2D), then geothermal is no longer able to compete.In our analysis, when all technologies experience substantial cost reductions (LowCost), primarily wind, solar, and biopower, but also offshore wind and battery storage, dominate systems under deep emission reduction constraints.System costs under the same emission reduction constraints in LowCost are substantially lower (on average 38% lower for >90% emission reduction constraints) compared to system costs in HighCost.Biopower use is greatly reduced in HighCost_LowGeo and HighCost_LowWindGeo due to the reduced use of variable wind and solar generation.Biopower use is also reduced in LowCost because wind and solar are less costly, and this results in more curtailments (Figure 3).

System cost distributions among ensemble members
In this section, we analyze the least-cost optimizations for all possible cost combinations at 2050 cost levels.
Figure 4 shows the distribution of simulated costs and installed nameplate capacities for different technologies and cost projection levels, under the 100% emission reduction constraint (results under the 99% emission reduction constraint are plotted in Figure S7).In general, cost reductions for technologies with multiple cost projection levels lead to more penetration and installed capacity of that technology.Geothermal shows up in optimal systems only when its low future cost projection is achieved with its degree of deployment being very sensitive to the cost of wind power (Figure S8).In our simulations, cost reduction in each technology tend to increase installed capacities of that technology, but the total amount of spent on deploying that technology (in absolute terms as plotted in Figure 4 or percentages to total system costs listed in Tables S2  and S3) does not in general increase for all technologies.For example, under the 100% emission reduction constraint, contributions of solar to total system costs range from 5.6% to 19.0% under the high-cost projection, while the numbers range from 3.9% to 23.3% under the low-cost projection.Similarly, contributions of battery storage to total system costs range from 7.2% to 11.1% under the high-cost projection, while the numbers range from 4.4% to 22.3% under the low-cost projection.Figure S9 shows that solar exhibits a median price elasticity of demand (here we refer to the input fixed costs as ''price'' and total capacity installed as ''demand'' for each technology) that is less than 1.0 -for each percentage decrease in the cost of solar there is a smaller percentage increase in the solar capacity deployed.In contrast, onshore and offshore wind, and battery storage exhibit a median price elasticity of demand that is close or greater than 1.0.
Figure 5 shows the distribution of ensemble members under the 100% emission reduction constraint, that is, the ratio of cases under different system costs and mean electricity dispatch levels to total ensemble members (i.e., the number of selected cases divided by the total number of cases), as a function of technology types and cost projection levels.Results under the 99% emission reduction constraint are plotted in Figure S10.In our simulations, cost reductions in solar and offshore wind have limited impact on the distribution of system costs, with lower cost projections slightly increase their capacities in the optimized solutions.In contrast, cost reductions in wind and geothermal not only decrease overall system costs, but also reduce the uncertainty of system cost distributions (i.e., narrowing the width of the pattern).For example, system costs differ by as much as 30% when the wind cost follows its high future cost projection, and only 22% at its low future cost projection; system costs differ by 46% when the geothermal cost follows its high future cost projection, and only 22% at its low future cost projection.This occurs because the substantial cost reductions in these technologies result in a dominant role of either technology, which substantially increases their capacities and dispatches in the systems and leaves limited space for other technologies to compete.As a result, the costs of minor competing technologies have little influence on system costs.
Figure S11 further examines how important cost reductions in each technology are in affecting overall system cost distribution.Across all 243 combinations for the year 2050 under the 100% emission reduction constraint, cost reductions in wind and geothermal have the largest impact on expected system cost.In our idealized analysis, the least-cost system ($0.06/kWh)among all possible solutions (as gas and gas-with-CCS are excluded) is comprised of all technologies (except nuclear and geothermal) at their lowest cost levels (as described previously, biopower and nuclear have only one cost projection).Geothermal does not compete in the least-cost condition, but there are systems with costs no more than 5% greater than the least-cost solution, in which geothermal competes at its low future cost projection.We also observe systems with costs within 10% of the least-cost solution, that include solar, offshore wind, and battery storage, competing at their middle cost projections.Similarly, there are systems costing no more than 20% over the least-cost solution where solar competes at its high, and wind at its middle cost projections.With the exception of wind and geothermal, decrease in the cost of one technology does not greatly reduce the mean or the variation of system cost (Figure S11).That is, if we get any single one of these technologies to be low cost, we could still end up with a costly system if we fail to lower costs in other technologies.Meanwhile, system costs would be moderate knowing that wind or geothermal would be on their low-cost levels, even if we fail to lower costs of other technologies.In these cases, substantial amounts of cheap wind and/or geothermal capacities are deployed, ensuring moderate to low system costs.
In our main scenarios, we consider unlimited resource potential for all technologies, while the US Department of Energy Geothermal Technologies Office reported $530 GW as the total technical potential for electricity generation for all types of geothermal.This accounts for approximately 70% of the 2050 average hourly demand according to estimations from the NREL Electrification Future Study. 60,61To reflect such limitation on geothermal, we reanalyze the 100% emission reduction constraint cases and restrict the maximum capacity of geothermal  and (E-H) 100% emission reduction constraints.Rows correspond to the four panels in Figure 2. Wind and solar consistently serve as lower-cost options for modest emission reduction constraints when fossil fuels are available to provide system reliability, whereas a combination or wind, solar and/or geothermal can dominate under deep emission reduction constraints.
to certain levels (i.e., 0.7, 0.5, and 0.3 times the annual mean demand).Figure S12 shows that restricting the maximum geothermal capacity, in general, results in a reduced ratio of cases under lower system costs and increases the ratio under higher costs.Among all ensemble members, restricting the capacity of geothermal has the least impact on systems with wind at its low future cost projection (Figure S13).This is because cheaper wind power directly competes with geothermal to provide the bulk of electricity, and thus there is minimal need to build geothermal with wind at its low future cost projection.
Our main cases are single-year least-cost optimizations assuming complete information, in which technology costs and emission constraints are prescribed at specific levels, and the model considers an electricity system at a stationary state without pre-existing capacity.In the real world, technology costs would evolve at different rates, and the capacity of any technology built in earlier years would last for decades and have long-term impacts.We conduct stylized transient simulations starting from 2019 untill 2050 with yearly updated emission reduction constraints and technology cost inputs.Compared to the single year 2050 optimization results, stylized transient simulations to 2050 show higher system costs and a wider distribution of system costs (Figures S14 and S15).In these cases, there are more wind, solar, offshore wind, and battery storage, and less geothermal due to the persistence into the time of 100% emission reduction constraint of capital stock that was built under weaker emission constraints (Figure S16).Detailed discussions are included in Text S2. iScience Article

Expected projected system cost and uncertainty
If someone were to plan future electricity systems based on current expectations of potential technology cost reductions and only later were able to identify which of the cost estimates are correct, which system would be preferred?This approach is intended to be analogous to situations in which energy system decisions are made based on assumptions of future cost, which may or may not be correct.
To fully explore such question is beyond the scope of our analysis.However, we can still perform high-level assessments utilizing our idealized framework.To do this, we reframe the question as: if someone were to build these systems given some NREL 2021 ATB technology cost projections, and only later were told which of the cost projections are correct, which system would have the lowest expected cost and the lowest uncertainty of expected cost?In our analysis, projected system costs under given cost reduction projections can be calculated using ensemble simulations under the 100% emission reduction constraint from the previous section and the distribution of 2050 cost combinations.That is, we take the simulated technology capacities and dispatches from each ensemble member that already exists (i.e., systems that were built based on given cost assumptions, in total 243 ensemble members) and see how the system cost would change when giving all possible cost combinations (the actual system costs, in total 243 cost combinations for each ensemble member).If we make the assumption that all of the NREL 2021 ATB cost projections are equally likely, then the expected cost for each ensemble member or certain technology capacity mix is the mean, and uncertainty can be represented as one standard deviation of system cost across all 243 cost combinations. .Probability distribution of ensemble members X axis represents different bins of (A-E) system costs and (F-J) mean electricity dispatch levels, and y axis shows the ratio of the number of cases in each bin to total number of ensemble members under various technology cost levels.That is, each panel includes all ensemble members and each pattern (i.e., one of the advanced, moderate, and conservative innovation trajectories) has one-third of ensemble members.Ensemble members under the 100% emission reduction constraints (in total 243 cases).The distribution function is produced using a Gaussian kernel density estimation tool from the SciPy package in Python.Cost reductions in wind and geothermal not only decrease overall system costs, but also reduce the uncertainty of system cost distributions.
Figure 6 shows that systems (or technology capacity mixes) that show lower expected (i.e., average) projected costs across all 243 possible cost combinations also have smaller values of one standard deviation of projected cost.For technology mixes with zero geothermal capacity, wind contributes to the largest cost difference among cost combinations (Figure S17); the presence of geothermal substantially increases both expected costs and its uncertainties due to the wider range in its projected 2050 cost reduction potentials (ranging from $12% to 73%), and the fact that geothermal is cost-competitive at its low future cost projection.Our results thus indicate that approaches that minimize expected costs across the probability distribution of future actual costs are more likely to produce lower cost systems than will approaches built in the hope that some current-costly technology will achieve a cost breakthrough.If the cost breakthrough fails to materialize, the resulting systems could prove costly.Here, variations in the costs of geothermal and wind have the largest influence.

DISCUSSION
Here, we conduct an idealized and transparent analysis to explore the impact of uncertainty in technology cost projections on the results of least-cost optimizations of future decarbonized electricity systems.We have used future cost estimates from a single source to reduce the possibility of bias that might be introduced by using an ad hoc set of assumptions made by the author team.We have shown that with the level of future cost uncertainty as represented by the NREL ATB projections, it is not possible to predict with confidence which technologies will play a dominant role in a future least-cost carbon-emission-free energy system.
Our model takes into consideration techno-economic factors only, and the objective function is designed to minimize the system cost considering technology capacity and electricity dispatch.We compare simulations that isolate the impact of applying different technology cost inputs based on future projections from the NREL 2021 ATB dataset.Our study focuses on a fundamental understanding of system dynamics.While our numerical results should not be overinterpreted, the qualitative influence of future cost uncertainty on future least-cost electricity systems will likely prevail under more generalized conditions.
Compared to the near current cost levels (i.e., 2019 estimates), fixed cost associated with construction and fixed operation and management (O&M) decreases to 2050 for all technologies examined here, with cost reductions varying from less than 15% for technologies such as Uncertainty in actual system cost represented as one standard deviation ($/kWh) Expected system cost ($/kWh) Figure 6.Expected system cost and uncertainty in actual cost X axis represents the expected system cost, which is calculated as the mean value of all 243 possible costs for a certain capacity mix.Y axis represents the uncertainty in actual cost, which is calculated at one standard deviation from all 243 possible costs.Here, we consider optimal technology capacity mix using ensemble members under the 100% emission reduction constraint.Possible electricity systems with lower expected costs may have lower uncertainties in realized costs.
gas, biopower, nuclear, and geothermal under the conservative technology innovation trajectory to more than 60% for technologies such as onshore wind, solar, battery, and geothermal under the advanced technology innovation trajectory.Simulations considering a wide range of emission reduction constraints using different technology cost combinations indicate that optimized electricity systems are more similar under low to modest emission reduction constraints (e.g., <80% emission reduction constraint).Under these constraints, variable renewables, specifically solar and wind, are low-cost options preferred as low-carbon electricity generation, with flexible fossil fuels filling gaps at times of low variable renewable supply and/or high electricity demand.Under deep-decarbonization scenarios (e.g., >90% emission reduction constraint), however, either variable renewables (e.g., wind and solar) or firm generation technologies (i.e., geothermal) could dominate the least-cost system, depending on relative ratios of realized cost reduction.
Considering all possible cost combinations from the NREL ATB dataset under the deep emission reduction constraints (243 cases under 100% and 729 cases under 99% emission reduction constraints), the largest system cost when all technologies follow their high-cost projection levels is less than two times the least-cost solution when all technologies follow their low-cost projection levels, highlighting the benefits of using a portfolio of technologies.
Our study has shown that cost reductions of certain technologies do not guarantee an increased deployment share of those technologies, as the least-cost solution depends on cost reductions of competing and complementary technologies as well.For example, we have shown that solar and battery storage contribute to as low as 3.9% and 4.4% of total system costs under low-cost projections, smaller than that under high-cost projections (5.6% and 7.2%).We also find that different technologies exhibit different price elasticities of demand.For each percentage decrease in cost, there is a greater increase in capacity demand for technologies such as onshore wind than for technologies such as solar.Different price elasticities of demand in an energy sector might have important implications of the economic scale and growth pattern for different energy technology industries.
Among all technologies, knowing that onshore wind or geothermal would follow their low future cost projections would substantially reduce the range of simulated system costs.That is, cost reductions in both onshore wind and geothermal not only increase the probability of a less costly system, but such cost reductions also reduce the uncertainty of system cost distributions.This is because these technologies are becoming more competitive under lower cost projections than other technologies.Geothermal, with a wide range of future cost projections, shows up only in systems where its low-cost estimate is achieved, and it competes mostly with onshore wind.Meanwhile, cost reductions in technologies such as solar and offshore wind have very limited impact on the probability distribution of optimized system costs.
Placing too much trust in unproven technologies can be risky.Considering situations in which energy system decisions are made based on assumptions of future cost that may or may not be correct, our results indicate that electricity systems with lower expected costs also tend to have lower uncertainties in actual cost.Systems built in the hope that some current costly technology will achieve a cost breakthrough have greater uncertainties.For example, we have shown that deep cost reductions of geothermal power could motivate its widespread use in a least-cost system, which strongly motivates efforts to reduce the costs of geothermal electricity generation.However, if wind power also achieves substantial cost reductions, there would be less motivation to deploy geothermal power.And if electricity storage and solar costs were also to decline dramatically, even low-cost geothermal power might not be able to compete.Developers can engage in research and development (R&D) to try to reduce the costs of a particular technology, but that technology's success is contingent on the cost of competing technologies.Building a large amount of geothermal power in anticipation of future cost reductions that do not materialize, can result in relatively high system costs.
These considerations suggest a strategy to deploy whatever clean technologies are economically viable today while engaging in actions that create and maintain a diverse set of options for the future, such as a research and development activities aimed at reducing tomorrow's costs across a broad spectrum of energy technologies.

Limitations of the study
In this study, we use a linear optimization electricity system model, the MEM, to examine the potential impact of applying different technology cost estimates on simulations of future decarbonized electricity systems.Instead of capturing the realistic market status or technical details, MEM is constructed in an idealized way to facilitate a fundamental understanding of electricity system dynamics in the context of carbon mitigation.The simplicity of MEM allows us to run hundreds or thousands of cases and examine a broad range of cost combination scenarios, which is difficult for models with complex structure and sophisticated underlying assumptions.The model can be easily extended to include more generation and storage technologies, constraints, and even change the objective function to fit other purposes.In this study, we isolate the impact of uncertain technology cost estimates in the future by changing only cost inputs, while keeping other factors unchanged, such as the model configuration, discount rate, hourly electricity demand, and wind and solar generation potential profiles.We provide an idealized framework and transparent analysis with full details disclosed.Our study does not serve to directly inform decision-making or design of the power system, but we hope to inspire future discussions and potential detailed works considering advanced models and sophisticated assumptions.Caveats should be kept in mind when interpreting results presented in this analysis.
Our quantitative results are contingent upon the assumed probability distribution of future technology costs.No one can generate probability distributions of future technology costs that will be universally accepted.To illustrate the potential influence of different technology cost assumptions, while avoiding our own appraisal of probability distributions of future costs, we use the NREL 2021 ATB estimated levels for advanced, moderate, and conservative cost innovation trajectories, and assign each of these trajectories with equal probability.Of course, these cost trajectories might not be equally likely achievable in today's view.For example, the advanced technology innovation scenario pathway assumes substantial technology innovations and cost reductions associated with geothermal (e.g., substantial drilling advancements), which might need great policy support and investments.Also, the conservative technology innovation scenario assumes lower levels of R&D investment with minimal technology advancement and cost reduction for wind and solar, inconsistent with current trends.Future works using different weighting approaches are encouraged.Considering equal probability also enables us to examine the distribution of system responses under various technology cost levels.
Cost trajectories taken from the NREL 2021 ATB dataset do not cover the full range of future technology cost distribution potential.For example, three technologies (i.e., gas, nuclear, and biopower) in our analysis are assigned with only one future cost trajectory while three trajectories are considered for other technologies.A previous report has found that a nuclear power plant with an advanced reactor design could potentially reach a capital cost as low as $4000/kW, 58 the impact of which can be partially represented by the low-cost geothermal trajectory.A recent work, 25 based on the probabilistic cost forecasts approach, has shown that technologies such as solar, wind, and battery might reach lower cost levels than values from the advanced innovation pathways.To reflect such possibilities, we have conducted additional cases where costs for solar, wind, and battery are 30%, 50%, and 70% of the LowCost levels.These lower cost assumptions further enlarge benefits of the variable renewable technologies, leading to greater dispatches from these technologies and system cost reductions (Figures S18 and S19).However, as long as the high-cost projections stay possible for these technologies, qualitative conclusions from our analysis, such as that both variable renewables and firm technologies could dominate the electricity system, and that electricity systems with lower expected costs may have lower uncertainties in realized costs, remain robust.Regional differences in technology cost reduction potential, resource availability, and electricity demand 32,62 would also influence optimization results.In addition, our least-cost solutions do not consider many practical issues faced by system planners.For example, environmental and social considerations may lead to the selection of one technology over another (e.g., land-use, visual and acoustic pollution, transmission rights-of-way, and safety concerns).
4][65][66] Our analysis uses future technology cost estimates under various scenarios at annual time step taken directly from the NREL ATB dataset. 37These scenarios provided by the NREL ATB dataset are established based on assumptions including different learning rates, deployment scales, and innovation levels.Therefore, learning effect is implicitly considered but not explicitly modeled.It should be noted that projections based on experience curves embody substantial uncertainty, 25,67 with ranges wider than that represented in this analysis.Such degree of uncertainty in the experience-curve-based cost projections likely precludes confident prediction of the composition of least-cost electricity systems for 2050.Meanwhile, learning by doing is not the only factor that determines future cost changes.Previous study has shown that the largest share of cost reduction in lithium-ion batteries was driven by public and private research and development 68 ; another example would be the cost reduction potential for advanced nuclear plant mentioned above 58 ; Cost saving innovations could lead to enhanced adoptions of new technologies that are currently expensive and immature. 66Our analysis is designed to isolate the impact of cost uncertainty and stay traceable for transparency.Future work should focus on better understanding how uncertainty in experience-curve-based cost projections translates into uncertainty in least-cost energy system transitions 65,69,70, .
Our main cases represent least-cost optimizations considering conditions in a single year treated as a quasi-steady-state, with both fixed and variable costs prescribed.2][73] We conducted idealized transient simulations and found that in cases with gradually decreasing emission constraints, it could be costly to fully eliminate fossil fuels from the system.Transient simulations also end up with more wind and solar and less firm technologies at the year 2050, given the long capacity lifetime.
Our simulations use a constant discount rate of 7% in determining the cost of capital recovery.Preferred values for discount rates could vary among different regions and technologies. 74Higher discount rates tend to favor technologies such as gas, with lower fixed costs and higher variable costs.Since our discussions focus on deep emission reduction scenarios (e.g., 100% emission reduction), in which most technologies excluding biopower are dominated by capital costs rather than operating costs, changes in discount rate will have only minor impact on the cost ratio among technologies or our qualitative conclusions.For instance, ratios of cost between solar and geothermal under a 3% or 15% discount rate differ by less than 0.3% compared to the default discount rate case (7%) for all three cost trajectories.
Our analysis considers a limited number of technologies.There are other possible low-carbon emission choices that can provide additional values to the system. 22,23,47,75For example, we did not include long-duration storage in our main cases since its future cost projections are not directly available from the ATB dataset.Considering that adding additional flexibility on seasonal scale could potentially benefits variable renewables more than firm technologies, we have added simulations with long-duration storage represented by a hydrogen storage system, which has three separate parts in the model: the electrolyzer produces hydrogen (H 2 ) using electricity, hydrogen storage stores the produced H 2 underground, and the fuel cell converts H 2 back to electricity.We scale the fixed cost of different components of the hydrogen storage system that represents a wide range of future cost reduction potentials.We also consider cases with an idealized representation of direct air capture of CO 2 .We add these technologies to the representative 2050 cost scenarios, respectively, and consider the 100% emission reduction constraint.Our results suggest that (Figure S20) low-cost hydrogen storage systems could benefit wind and solar by providing additional systematic flexibility with most benefits coming from the fuel cell component.In deeply decarbonized scenarios, low-cost hydrogen storage systems would also diminish the value of nuclear and geothermal.Reducing the cost of direct air capture enables fossil fuel generations under deep emission reduction constraints, and benefits wind and solar (Figure S21).As the cost of air capture approaches zero, reducing direct air capture costs in our model shifts the system to the one dominated by gas.
Our simulations use cost projections provided in the 2021 ATB dataset, which was updated later by NREL.A more recent version (2022 ATB dataset version 3, accessed in March, 2023) shows consistent cost trajectories for most technologies, with more cost reduction potential for geothermal under the advanced technology innovation trajectory (Figure S22).Taking into consideration the more recent numbers would make geothermal more competitive compared to other technologies when following its low-cost projection.Meanwhile, since the cost reduction potential for geothermal is modest under the conservative technology innovation, geothermal would remain uncompetitive when following its high-cost projection.Therefore, using the more recent ATB cost projections will not affect our qualitative conclusions.
In our main cases, we use the same electricity demand, originally derived from EIA, 76 and combine it with concurrent generation profiles and different technology cost combinations.In reality, electrification of other sectors (e.g., transportation, industry, and construction) would change the shape and increase the variability of the demand profile (since we normalize the annual mean demand to 1 kWh, increases in magnitude of annual demand do not affect our conclusions here).To assess impact from a more variant electricity demand profile, we consider two additional scenarios: one with a demand profile calculated from the NREL Electrification Future Study (EFS) report 60,61 ; and the other scenario using an idealized demand profile, which is calculated as the square of the default 2019 demand profile for each hour (E inew = E 2 i old ) from EIA.Both scenarios lead to slightly more use of variable renewables under the representative 2050 scenarios and larger system costs when approaching 100% emission reduction, while they do not change the qualitative conclusions of our analysis (Figure S23).Our approach still applies a static demand profile, which is not allowed to shift in time to reflect part of the advantage from electrification (e.g., demand management through electric vehicle charging).Allowing the demand to shift in time with low marginal cost is expected to promote renewables with cheaper battery storage and long-duration storage.
Many other considerations are important in making real-world decisions on investment strategy for a portfolio of technologies. 64Some of them are beyond the scope of this analysis and others are difficult to assess under our idealized framework.A few examples include that our analysis uses hourly time steps, whereas reliability of power systems can depend on grid events occurring on time scales of milliseconds to minutes. 77Similarly, we do not consider many other operational constraints that might be crucial to ensure the reliable operation of a realistic power system (e.g., ramping constraint of the nuclear power plant).There are uncertainties other than cost inputs when modeling energy systems, and these uncertain factors might interact with each other and produce unanticipated outcomes. 40,53The optimization of our least-cost systems depends heavily on given cost inputs and other assumptions (e.g., carbon emission constraint and capacity factor profiles), while in reality there are many other factors that would incentivize or disincentivize the deployment of certain technology.During the optimization, MEM has perfect foresight of future electricity needs and wind and solar resource availability, and assumes free and lossless conduction of electricity within the simulated region.These assumptions confer advantages on variable renewables relative to more realistic representations.Avoidance of fossil CO 2 emissions would need to occur across all sectors to fully decarbonize the whole society.Some sectors that depend on high temperatures (e.g., steel, cement), high amounts of power (e.g., space heating), or high energy density (e.g., aviation) may prove more difficult to mitigate. 8Zero-emission technologies such as the concentrated solar power 24 that can provide both electricity and thermal needs, and hydrogen that can be combusted to provide electricity, serve as long-term storage, and be used as liquid fuel might play important roles in future energy systems.Our model represents only the electricity sector as a single-node system, which does not consider the impacts of interconnections among multiple regions and the bridge between the electricity system and the non-electrified parts of the energy system. 78,79attery storage (Figure 1).We do not include coal because coal is more expensive and produces more carbon emissions than gas in the ATB dataset, and thus it will never compete with gas in the least-cost solution.Electricity inputs from the above technologies are used to meet demand, or have been curtailed at each hourly time step to maintain balance of the electricity exchange node at no cost.There is no lost load allowed during the simulations, and 100% of electricity demand is forced to be satisfied.
In current configuration of MEM, a fixed cost component is associated with all technologies to represent the fixed capital investment including the purchase and installation costs, and fixed operation & management (O&M) costs, and a variable cost is specified for gas, gas-with-CCS, nuclear, and biopower that includes variable O&M and fuel costs as appropriate (Table S1).For nuclear, we assume that the nuclear reactor must be operated at constant rates, and thus operation and fuel costs (i.e., variable costs) are added to fixed cost as dollar per unit capacity.The original cost assumptions for all generation and storage technologies, including the capital cost, fixed O&M expenses, variable operation and maintenance expenses, and fuel costs, are taken directly from the NREL 2021 ATB dataset. 37We then convert these cost values to fixed and variable cost components used in MEM considering the same discount rate (7% per year) in the same manners.Note that the solar cost taken directly from the NREL ATB dataset is in units of $/kW-AC and we do not consider factors such as the inverter loading.This might lead to slightly larger numbers as model inputs, while the difference is much smaller compared to model inputs under various cost projection trajectories.
The NREL ATB dataset provides cost estimates for different technologies for three cost reduction pathways covering periods from year 2019 to year 2050.These cost reduction scenarios are named Advanced, Moderate, and Conservative Technology Innovation pathways, and are defined based on different assumptions of learning rates, deployment scales, and innovation levels.Costs under these three pathways are more consistent in near-current times (i.e., same in year 2019) and diverge from each other with time, leading to substantially different cost projections by year 2050.Our analysis considers the year-2019 scenario to represent the near-current cost levels, and three year-2050 cost levels representative of Advanced, Moderate, and Conservative Technology Innovation pathways, corresponding to low, middle, and high future cost projections.We do not make our projections of future technology costs.For gas, nuclear, and biopower that have only one future cost scenario in the 2021 ATB dataset, we use that single cost level in our simulations.For offshore wind, we choose the Class 8 floating wind turbine technology, which is cheaper than the Class 12 wind turbine technology that represents NREL's most recent assessment of the resource characteristics of mid-term deployment for floating technology in the California Call Areas defined by the Bureau of Ocean Energy Management (BOEM).For geothermal, we consider the flash enhanced geothermal system (EGS) to reflect larger geothermal resource potential in the United States.Such approach is used to reduce the overall numbers of ensemble members and to narrow down the scope of the analysis.The derived costs to deploy 1 kW nameplate capacity for solar and wind are cheaper compared to that for gas plant.However, gas remains as the cheaper option under zero emission reduction constraint due to renewables' smaller capacity factors and higher variabilities.We do not pretend to know which cost projections (within or outside the range of NREL's estimates) are more likely to happen in the future, and thus we treat each cost level equally during the simulations.MEM determines both capacity and electricity dispatch at each hourly time step based on ratios of technology costs, instead of the absolute cost values.That is, if all technologies experience the same percentage of cost reductions (both fixed and variable costs) in year-2050 compared to year-2019, the optimized technology capacities and hourly dispatches would be the same despite decreases in system cost.
In our analysis, we first run simulations under the year-2019 cost estimates, and then run four representative year-2050 cost combination cases: HighCost, where all technologies (excluding gas, nuclear, and biopower, each of which has only one future cost projection) follow the conservative technology innovation trajectories and achieve high future cost projections; HighCost_LowGeo, where only geothermal follows the advanced technology innovation trajectory and achieves the low cost projection, while other technologies remain at high cost projections; HighCost_LowWindGeo, where both wind and geothermal follow the advanced technology innovation trajectories and are at their low cost projection levels, while other technologies remain at high costs projections; and LowCost, where all technologies achieve the advanced innovation pathways and achieve their low cost projections.We run these simulations under multiple emission reduction constraints, in which CO 2 emissions from fossil fuel-based generation sources (i.e., gas and gas-with-CCS) are gradually eliminated.Our analysis focuses on restricting CO 2 emissions from gas and gas-with-CCS since they have much more direct carbon emissions than those of other low-carbon emission sources used here. 1The upper boundary limits of CO 2 emissions from gas and gas-with-CCS vary from 100% to 0% with an increment of 2%, plus additional deeply decarbonized scenarios-99%, 99.9%, and 99.999%-relative to the case, in which all demands are met by gas.Changes in the Stack-plot would be smoother if a higher granularity in emission reduction constraint (i.e., < 2%) were used.We then perform year-2050 simulations considering all possible cost combinations under the 100% emission reduction constraint.Since there is no carbon removal approach in the main cases, the full decarbonization scenarios represent an ''absolute zero'' carbon emission condition (not accounting for cradle-to-gate emissions of generating technologies).To examine the potential temporal lock-in effect, we conduct idealized transient simulations with time-evolving emission reduction constraints and technology costs.The least-cost solution is found using the Gurobi optimizer written in Python.

Model formulations
We summarize the complete model formulation and nomenclature in this section.
In MEM, fixed costs of generation and storage technologies (c fixed ) are calculated as: c fixed = gc capital +c fixed O&M h (Equation 1)

Figure 2 .
Figure 2. Contributions to total system costs Four illustrative cases under 2050 cost levels and various emission reduction constraints are compared.(A)HighCost all technologies (excluding gas, nuclear, and biopower, which have only one future cost projection) follow the high future cost projections.(B) HighCost_LowGeo where only geothermal reaches its low cost projection while others remain at high costs projections.(C) HighCost_LowWindGeo where both wind and geothermal reach their low cost projections; and (D) LowCost where all technologies reach their low future cost projections.The same 2019 demand and generation potential profiles are used to isolate the impact caused by applying different technology costs.Geothermal plays a major role if it is the only technology to reach its low-cost level, but if other technologies reach their low-cost levels, the role of geothermal is reduced or eliminated.

Figure 3 .
Figure 3. Daily and hourly electricity dispatches Four 2050 cost scenarios with different cost combinations are selected (HighCost, HighCost_LowGeo, HighCost_LowWindGeo, and LowCost) under (A-D) 80%and (E-H) 100% emission reduction constraints.Rows correspond to the four panels in Figure2.Wind and solar consistently serve as lower-cost options for modest emission reduction constraints when fossil fuels are available to provide system reliability, whereas a combination or wind, solar and/or geothermal can dominate under deep emission reduction constraints.

Figure 4 .
Figure 4. Distributions of simulated cost allocated to each technology and capacities Results of (A) system cost attributed to different technologies and (B) installed nameplate capacity (i.e., capacity associated with fixed cost; actual generation potential equals installed nameplate capacity times capacity factor for wind and solar on each time step) are plotted for each technology under various 2050 projected cost levels using the Matplotlib function Boxplot.The boxes extend from the first quartile to the third quartile of the data, with a line at the median.Whiskers extend from the box to the farthest data point lying within 1.5x the inter-quartile range from the box.In our linear-optimization framework, system cost can be represented as the linear sum of costs attributed to different technologies, including the fixed cost of installed capacity (as shown in panel B) plus non-zero variable cost of total electricity dispatched (e.g., gas and biopower).Cases under the 100% emission reduction constraint are shown, and results are expressed in terms of cost or power per kWh of annual mean demand.For technologies with three cost projections in NREL ATB, there are 81 combinations of costs of the other technologies under each cost level.Biopower and nuclear have only one cost projection.Geothermal has non-zero capacity only when it reaches its lowest cost level.Cost reductions in a technology typically result in an expectation of increases in deployed capacity of that technology, but not necessarily an expectation of an increased spending on that technology (e.g., solar).

Figure 5
Figure 5. Probability distribution of ensemble members X axis represents different bins of (A-E) system costs and (F-J) mean electricity dispatch levels, and y axis shows the ratio of the number of cases in each bin to total number of ensemble members under various technology cost levels.That is, each panel includes all ensemble members and each pattern (i.e., one of the advanced, moderate, and conservative innovation trajectories) has one-third of ensemble members.Ensemble members under the 100% emission reduction constraints (in total 243 cases).The distribution function is produced using a Gaussian kernel density estimation tool from the SciPy package in Python.Cost reductions in wind and geothermal not only decrease overall system costs, but also reduce the uncertainty of system cost distributions.

Electricity exchange node Hourly input = Hourly output Macro Energy Model (MEM)
BatteryFigure 1. Diagram of the linear-optimization electricity system model used in this analysis Wind and solar fields represent variability in the contiguous US.Cost assumptions are from the National Renewable Energy Laboratory (NREL) 2021 Annual Technology Baseline (ATB) dataset.