Cost Evolution Mechanisms of Renewable Energy Technologies: Onshore Wind Power and Photovoltaics in China

Abstract

The unit costs of power generation of onshore wind and photovoltaics in China have dropped rapidly and significantly since 2010. Recent studies have indicated that the learning effect on cost reduction could have been overestimated due to the exclusion of the equipment-level installed capacity and the price of capital. To address this estimation bias, we constructed a research framework comprising a one-factor analysis model (OFAM), a two-factor analysis model (OFAM), and a multi-factor analysis model (MFAM) based on the Cobb–Douglas function and the cost minimization problem. This framework examines the determinants of unit costs in renewable energy generation in consideration of learning effects, scale effects, and price effects. This paper uses data from institutions such as IRENA and the World Bank to empirically analyze the contributions of these factors to reductions in the cost of onshore wind and photovoltaic power generation in China from 2010 to 2022. The results indicate that the learning-by-doing (LBD) effect has been overestimated, with scale effects accounting for a significant portion of the cost reduction. Moreover, the price of capital exerts a more pronounced influence on the levelized cost of electricity (LCOE) for photovoltaics. After factoring in equipment scale and capital costs, LBD continues to significantly reduce the LCOE of photovoltaics, with the LBD learning rate declining from 23.85% to 6.30%. Meanwhile, the impact of LBD on the LCOE of onshore wind technology ceases to be significant. Both technologies exhibit economies of scale, with scale effects accounting for 41.60% and 34.12% of the LCOE reductions for onshore wind and photovoltaics, respectively. Capital costs accounted for 32.50% of the LCOE reduction for photovoltaics. Therefore, future large-scale deployments of other costly renewable energy technologies may also benefit from the equipment-level scale and favorable bank interest rates in addition to learning-by-doing.

1. Introduction

The International Energy Agency [1] predicts that global renewable energy generation capacity will more than double by 2050 compared to 2022, with nearly 95% of this capacity coming from solar photovoltaics and wind power. China’s total installed capacity for renewable energy power generation accounted for nearly 40% of the global total in 2023 [2]. The installed capacity of wind and solar photovoltaic power generation technologies is growing rapidly, with China’s installed wind and solar power capacity exceeding 1.4 billion kilowatts by 2024 [3]. According to China’s 14th Five-Year Plan for Renewable Energy Development [4], the installed capacity of wind power and photovoltaics is expected to continue to increase rapidly.

The levelized cost of electricity (LCOE) is frequently used in the cost analysis of power generation technology [5,6,7]. The LCOE of China’s onshore wind has been below the global average and has continued to decline since 2010, with 2022 values dropping to 31.14% of 2010 values [8], while solar photovoltaic LCOE decreased to 11.23% over the same period [8], as shown in Figure 1.

The decline in the cost of renewable energy generation technology may be attributed to a combination of learning effects, scale effects, and price effects [6,7,9,10,11,12,13,14,15]. Among the multiple influencing factors, the learning effect is frequently regarded as the most significant influencing factor in cost analysis. Increases in cumulative installed capacity lead to increases in relevant experience, which in turn reduce costs through increased efficiency. This process is called learning-by-doing (LBD), a type of learning effect [11].

Recent studies have indicated that the previous literature may have overestimated the learning effect due to the exclusion of the equipment-level installed capacity and price effect [16,17]. Modifications in the scale of individual equipment exert an influence on power generation costs. The augmentation of power generation equipment capacity may be associated with the decline in unit installed costs or unit operating costs, also. That is to say, the cost per unit of power generation may decrease as the average capacity of a company’s power generation equipment increases. It can be regarded as the equipment-level scale effect [17,18]. Scale effects at the equipment level may also positively impact LCOE reduction by increasing equipment capacity. Some studies also pointed out that the augmentation of equipment capacity may exert a multifaceted, two-way influence on power generation cost, by giving rise to challenges such as escalated mechanical intricacy [19,20].

Additionally, the cost of utilizing capital may exert a substantial influence on production costs, as renewable energy projects are often capital-intensive, and the cost of capital utilization significantly affects the economic viability and feasibility of power generation projects. According to the extant literature, an increase in interest rates may exert an adverse effect on LCOE [16,21]. Fluctuations in the prices of input materials and capital exert a significant influence on power generation project costs [7], referred to as the price effect [22].

To respond to the inadequate understanding of the mechanism of cost reduction for renewable energy power generation over time, we attempt to formulate a framework that illustrates how multiple factors influence the unit cost of renewable energy power generation by incorporating capital prices and equipment-level capacity. Moreover, we endeavor to use the framework to examine the factors contributing to the decrease in the costs of wind power and photovoltaic power generation in China between 2010 and 2022.

This study makes two contributions. First, we construct a comprehensive analytical framework for examining the combined effects of learning, scale, and price effects on power generation costs. By integrating the multiple cost drivers previously discussed in isolation across the literature into a unified model for quantitative decomposition, we reveal the underlying mechanisms of cost evolution in a more comprehensive and multidimensional manner. Second, while the existing literature on renewable energy technology costs has primarily focused on learning-by-doing mechanisms, our study fully accounts for the critical impacts of equipment scaling trends and capital usage costs in technological development. Building upon the learning-by-doing framework, we innovatively incorporate equipment scales and capital prices as core variables for scale effects and price effects. Specifically, we identify the significant contributions of equipment scales and capital prices to the decline in LCOE for China’s onshore wind and photovoltaic power generation, revealing their unique operational mechanisms within China—the world’s largest renewable energy market.

The paper is organized as follows: Section 2 outlines the mechanisms driving cost reduction; Section 3 introduces the learning curve model and data sources; Section 4 presents the regression results, followed by an analysis and discussion; and Section 5 summarizes the key findings and discusses the limitations of the study.

2. Literature Review

2.1. Learning Effect

Numerous scholars have explored various learning mechanisms since the concept of the learning curve was first proposed [23,24], including learning-by-doing, learning-by-researching, learning-by-using, and learning-by-interacting [22]. Arrow [23] proposed the concept of learning-by-doing and attributed the decline in the cost of technology to the learning effect (accumulation of experience). Learning-by-doing is considered the most crucial mechanism in the learning curve, as it involves gaining experience through repeated operations within the same production process, leading to improvements in production efficiency and processes [25]. The relevant research shows that the learning effect allows for an increase in cumulative installed capacity to drive a reduction in power generation costs [5,26]. Some scholars have proposed that learning-by-researching also significantly impacts cost reduction [6,10,24,27], with research describing the cost reduction achieved through R&D to improve technology [24]. Learning-by-using involves feedback from product or service users that helps improve a company’s products. Learning-by-interacting refers to the process by which companies communicate with other firms or suppliers to enhance their products. Due to the limited availability of data related to learning-by-using and learning-by-interacting, this study does not consider these mechanisms in its analysis.

Most existing research on China’s renewable energy power generation technologies has focused on the driving factors of wind power cost changes, with only a few studies addressing photovoltaic costs. Moreover, some of this research has prioritized studying photovoltaic modules over photovoltaics [28]. For example, Zhang et al. [29] analyzed the cost change in polycrystalline silicon photovoltaic modules, finding a learning-by-doing rate of 14%. Learning-by-doing can contribute to the decrease in the cost of photovoltaic modules. Dai et al. [5] analyzed the learning rates of onshore wind power and photovoltaics in China and found that learning-by-researching had no significant effect on the decline in photovoltaic costs, whereas learning-by-doing did. In contrast, Chen and Han [30] found that both learning-by-doing and learning-by-researching significantly reduced unit installed costs. Lam et al. [31] found that the LCOE of wind power decreases by 3.5–4.5% when the installed capacity is doubled. The learning effect of learning-by-doing is a significant driver of technology cost reductions and can build advantages for the mass deployment of technology.

2.2. Scale Effect

The augmentation of equipment, projects, and production capacity has been demonstrated to engender the scale effect [32]. Isoard and Soria [11] considered the impact of the scale effect on unit installed costs and found that the scale effect increases unit installed costs and that the learning effect is underestimated if the scale effect is not considered.

In studies on the scale effect in the field of renewable energy power generation technology, the capacity of a power generation facility is often used as the scale variable at the equipment level [26,33,34]. The scale variable for onshore wind power is represented by the capacity of wind turbines, while the scale variable for photovoltaics is expressed as the capacity of photovoltaic modules. Coulomb and Neuhoff [35] decomposed the cost of wind turbines and found that the learning rate increased from 11% to 13% after controlling for the scale effect at the equipment level. Several scholars have also conducted empirical analyses of European wind power technology and confirmed that the scale effect at the equipment level significantly influences the cost of wind power in Europe [10,17]. The industry-level effect has been found to significantly reduce the cost of renewable energy technologies [12,14,36]. Scale effects exert a considerable influence on fluctuations in power generation costs; thus, a comprehensive consideration of scale effects is therefore essential for comprehending the cost reductions precipitated by expansion.

2.3. Price Effect

The price of input elements has an important impact on the cost of renewable energy power generation [6,22]. The impact of fluctuations in input factor prices on the cost of power generation through renewable energy is defined in this paper as the price effect. Söderholm and Sundqvist [14] critically analyzed the learning curve of European wind power and proposed that the estimated learning rate is biased when the significant influence of input prices on technological costs is not considered.

Kahouli [12] established a multi-factor analysis model (MFAM) model combining learning-by-doing, learning-by-researching, and material price in a study on the price of input materials, revealing that uranium prices significantly impact the cost of nuclear power generation and that incorporating uranium price variables enhances the accuracy of cost estimates. Similarly, silicon prices have been found to play a crucial role in the cost structure of photovoltaic systems: Decreases in silicon prices lead to reductions in total costs [15]. The cost of photovoltaic modules constitutes a core component of the total installed cost of photovoltaic systems, with polysilicon being the primary material used in photovoltaic modules [37]. Therefore, this study considers polysilicon prices to be the key material price indicator of photovoltaic technology. Similarly, wind turbine costs constitute the largest component of the total installed expenses for onshore wind power technology. Wind turbines are material-intensive and primarily utilize steel, aluminum, and fiberglass in their production processes [38], with steel accounting for 70% of the turbine’s total mass [39]. Consequently, this study designates steel prices as the core raw material price indicator for assessing onshore wind power technology costs. Research in China has explored the impact of other economic factors on the cost of power generation, such as capital investment and input material prices, but has ignored the role of capital prices. Hayashi et al. [40] analyzed the learning curve of China’s wind farms and concluded that the accumulation of experience and knowledge has no significant effect on the cost of power generation; instead, they found that power generation costs are primarily driven by capital investment. Intense price competition has been shown to reduce costs and opportunities for investment in technological improvements. Zhang et al. [9] established an MFAM that included cumulative experience, capacity factor, capital investment, and material price, believing that the capacity factor reflects the fluctuation of natural resources to a certain extent. They found that the reduction in power generation costs is strongly influenced by capital investment and material prices, with the efficient use of natural resources playing a critical role in the decline of LCOE. Price competition is the primary mechanism through which China reduces the LCOE for wind power, thereby achieving grid parity.

2.4. Summary

Scholars writing during the early years of research in the field of renewable energy usually considered only the impact of accumulated experience on cost changes [24]. However, many studies have shown that considering only one factor to explain the change in cost leads to deviations in the estimated learning rate [14,41,42,43]. Therefore, studies have built on the foundation of learning-by-doing and have further considered the impact of learning-by-researching, scale, and economic factors of labor and capital [6,33,44].

The capacity of power generation equipment for onshore wind power and photovoltaic technology in China has increased over time. The augmentation of power generation capacity enhances the efficiency of natural resource capture and reduces civil work costs per unit of power generation. However, this results in higher mechanical complexity and increased failure rates, which may lead to higher costs. Few studies have examined the impact of scale effects at the equipment level on the decline in the cost of onshore wind power and photovoltaic technology in China. The extant literature focuses on the price of materials and infrequently considers the cost of capital, usually expressed as the interest rate, which plays a pivotal role in determining financing costs for various power generation projects. Higher interest rates result in increased repayment amounts, thereby increasing the financing cost of the project. Consequently, this study introduces the scale effect at the equipment level and the price effect of capital.

3. Methods and Data

3.1. Cost Range

This study examines changes in the unit generation costs of onshore wind power and photovoltaic power generation in China. LCOE is equal to the net present value of the life cycle cost divided by the life cycle electricity generation [45] and includes the installed costs, operating and maintenance (O&M) costs, and fuel costs incurred over a project’s entire life cycle [46]. O&M costs encompass both fixed expenses—such as insurance, grid connection fees, administration, and regular equipment inspections—and variable costs arising from equipment repairs and operational consumption [38]. As most power generation costs are accounted for by unit installed costs, the literature considers the unit installed cost to be another important indicator of technology cost. Unit installed costs typically include equipment costs, civil work costs, and grid connection costs [47]. Fixed financing costs represent a percentage fee charged by financial institutions, such as banks, on the total financing amount arranged for a project’s debt financing [38]. This cost constitutes one of the development expenses for renewable energy projects and is included as part of the installed cost. The installed cost encompasses a narrower scope than LCOE, which incorporates O&M costs as well as the fuel costs incurred during a project’s operation, based on the installed cost [7].

Another indicator of the unit cost used in this study is the installed cost per capacity. Some studies have indicated that the unit installed cost is sensitive to cumulative experience, capital and material prices, and individual equipment scale under certain conditions [10,15]. The installed costs of wind power and solar photovoltaic power generation constitute a significant component of total power generation cost. The installed costs for onshore wind are responsible for above 70% of the total LCOE, while O&M costs account for less than 30% of the LCOE [38]. According to data from a report by the China Photovoltaic Industry Association (CPIA), the installed cost of a photovoltaic power plant with an operating period of 25 years accounts for approximately 80% of LCOE [48]. The specific cost components of renewable energy power generation technology are illustrated in Figure 2.

3.2. Model Construction

The logical structure of the model in this paper is shown in Figure 3.

3.2.1. Unit Cost Function

Building on the research of Yu et al. [15] and Yu et al. [49], this paper combines the Cobb–Douglas (C–D) production function, cost minimization problem to derive the cost function, and incorporates concepts from the learning curve model to establish a multi-factor theoretical model of power generation costs.

This paper refers to the work of Berndt [50] and rewrites the Cobb–Douglas function as follows:

$$Y=A{\prod }_i\left(x_i^{{\theta }_i}\right)$$

(1)

where $Y$ is the output, $A$ is the technological level, $x_i$ is the input element (generally expressed as capital, labor, materials, and the like in the C–D function), and ${\theta }_i$ is the input elasticity coefficient.

$$r={\sum }_i{\theta }_i$$

(2)

where $r$ is the return-to-scale parameter equal to the sum of the input elasticity coefficients.

Accordingly, the total cost function is expressed as follows:

$$C_{total}={\sum }_i{(x}_i{p}_i)$$

(3)

where $C_{total}$ is the total production cost, $p$ is the price of the input element, and $i$ is the number of input elements.

A generalized cost optimization problem can be expressed as follows to answer the question: What is the optimal selection of production process inputs to achieve the given output at the minimum total cost [15]. The Lagrange multiplier method can be used to construct the Lagrange function and solve this minimization problem:

$$\min \mathrm{L}\left(x_i,\lambda \right)={\sum }_i{(x}_i{p}_i)+\lambda \left[Y-A{\prod }_i\left(x_i^{{\theta }_i}\right)\right]$$

(4)

After obtaining the partial differential of Equation (4) and setting it to zero, we obtain Equations (5) and (6).

$${\displaystyle \frac{\partial L}{\partial x_i}}=p_i-\lambda {\displaystyle \frac{A{\theta }_i}{x_i}}{\prod }_i\left(x_i^{{\theta }_i}\right)=0$$

(5)

$${\displaystyle \frac{\partial L}{\partial \lambda }}=Y-A{\prod }_i\left(x_i^{{\theta }_i}\right)=0$$

(6)

We then transform the above equations. The functional expression for minimum total cost is shown in Equation (7) (see Appendix A for a detailed elaboration on the algebraic manipulation process used to derive these equations).

$$C_{total}=r{\left(A{\prod }_i{{\theta }_i}^{{\theta }_i}\right)}^{{\displaystyle \frac{-1}{r}}}{Y}^{{\displaystyle \frac{1}{r}}}{\left({\prod }_i{p_i}^{{\theta }_i}\right)}^{{\displaystyle \frac{1}{r}}}$$

(7)

Dividing Equation (7) by output Y yields the functional expression for unit cost in Equation (8):

$$C_{unit}=r{\left(A{\prod }_i{{\theta }_i}^{{\theta }_i}\right)}^{{\displaystyle \frac{-1}{r}}}{Y}^{{\displaystyle \frac{1-r}{r}}}{\left({\prod }_i{p_i}^{{\theta }_i}\right)}^{{\displaystyle \frac{1}{r}}}$$

(8)

The learning curve describes unit cost as a function of accumulated production experience. Specifically, it shows that as the output quantity doubles, unit cost decreases by a fixed proportion [51]. The learning curve model is widely used to analyze the causes of changes in technology costs and to predict future cost trends. In contrast to traditional engineering economic analysis, the learning curve model accounts for fluctuations in costs resulting from dynamic learning. Consequently, in contrast to static assumptions, which posit fixed and unchanging technical costs, the model can predict costs that are more aligned with reality. In addition, empirical research has shown that simulations employing learning curve models are often more proximate to actual costs than time series models [52].

Improvements in technological level lead to decreases in production costs; that is, unit cost decreases with an increase in $A$. We believe that cumulative production experience will reduce unit production costs by improving the technological level. Consequently, learning effects change $A$ in the Cobb–Douglas function.

Building on Berndt’s research [50], which considers $A$ to be positively correlated with cumulative production experience, this paper expresses $A$ as follows in Equation (9).

$$A={\mathrm{C}\mathrm{A}\mathrm{P}}^{-\alpha }$$

(9)

We let $n_0=r{\left[{\prod }_i{({\theta }_i}^{{\theta }_i})\right]}^{{\displaystyle \frac{-1}{r}}}$, and combine Equations (8) and (9) to obtain Equation (10).

$$C_t=n_0{\mathrm{C}\mathrm{A}\mathrm{P}}^{{\displaystyle \frac{\alpha }{r}}}{Y}^{{\displaystyle \frac{1-r}{r}}}{\left({\prod }_i{p_i}^{{\theta }_i}\right)}^{{\displaystyle \frac{1}{r}}}$$

(10)

where $C_t$ is the cost of producing a unit of product in $t$ year; $\mathrm{C}\mathrm{A}\mathrm{P}$ is the cumulative production from the first year to the $t$ year; $\alpha$ is the elasticity coefficient of learning-by-doing, which reflects the learning effect previously defined; and $Y$ is the production scale in $t$ year. The index ${\displaystyle \frac{1-r}{r}}$ in $Y$ reflects the impact of scale changes on unit costs. We define this impact as the scale effect. $p$ is the price of the input element, while ${\theta }_i$ is the elasticity coefficient of the input element price and reflects the impact of price changes on unit costs, which we define as the price effect. Lastly, $n_0$ is a constant term representing factors other than cumulative production, production scale, and prices.

The learning effect and the scale effect have the largest impact on unit costs among all learning effects [29,33,53]. In theory, these two effects operate through different mechanisms that drive cost changes. Figure 4 distinguishes the unique roles of the scale effect and learning effect. The curves L_AC and L_EF represent two long-run average unit cost curves. The two U-shaped long-run average unit cost curves indicate that alterations in economies of scale occur with increases in output. For both the L_AC and L_EF curves, within the light-colored interval in which point A (or point E) is located, relative changes in input and output exhibit increasing returns to scale, while relative changes in output and unit cost exhibit economies of scale. Relative changes in input and output within the transition zone where point B is located exhibit constant returns to scale, with unit costs remaining largely unchanged regardless of changes in output. Within the shaded region where point C (or point F) is located, relative changes in input and output exhibit diminishing returns to scale, while relative changes in output and unit cost illustrate diseconomies of scale. Consequently, movement along the long-run average unit cost curve reflects economies of scale, whereas learning effects occur at any level of output. At any given level of output, learning effects can lead to a decrease in unit production costs [11]. As illustrated in Figure 4, the unit costs at points A and C decrease to points E and F, respectively, thereby reflecting the learning effect. Consequently, the learning effect is a shift in the long-run average unit cost curve itself. In this study, the two parameters $\alpha$ and $r$ are related to the learning effect and scale effect, respectively.

When the price effect is not considered, Equation (10) can be simplified to Equation (11).

$$C_t=n_0{\mathrm{C}\mathrm{A}\mathrm{P}}^{{\displaystyle \frac{\alpha }{r}}}{Y}^{{\displaystyle \frac{1-r}{r}}}$$

(11)

When only the learning effect is considered, the return to scale is constant and regarded as 1 ($r=1$). Equation (10) can thus be simplified to the general formula of a one-factor analysis model (OFAM).

$$C_t=n_0{\mathrm{C}\mathrm{A}\mathrm{P}}^{{\displaystyle \frac{\alpha }{r}}}$$

(12)

The OFAM’s logarithmic form can be obtained by taking the logarithm of Equation (12).

$$\mathrm{l}\mathrm{n}{C}_t=\alpha \mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+\mathrm{l}\mathrm{n}{n}_0$$

(13)

3.2.2. Regression Model

To analyze the learning effect, the cumulative production is represented by the cumulative installed capacity, $\mathrm{C}\mathrm{A}\mathrm{P}$, from the first year to the $t$ year (see Equation (14)). $\alpha$ is a parameter related to learning effects. In Equation (15), LBD is defined as the learning rate for learning-by-doing, which represents the proportion of change in LCOE when the cumulative installed capacity is doubled.

$$\mathrm{l}\mathrm{n}\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}=\alpha \mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+\mathrm{l}\mathrm{n}{n}_0$$

(14)

$$\mathrm{L}\mathrm{B}\mathrm{D}=1-2^{\alpha }$$

(15)

Further, there is an impact of cumulative knowledge on the unit power generation costs [13]. Executing R&D endeavors can yield opportunities for the realization of technological advancements and reduce costs. This learning effect is called learning-by-researching [42]. The augmentation in installed capacity engenders not only learning effects but also economies of scale; that is, the cost per unit of power generation diminishes as the industry’s currently operating installed capacity increases. We can regard this economy of scale as an industry-level scale effect [12,18,36]. We screened multiple factors affecting renewable energy generation technology costs over time, in addition to the learning-by-doing effect, by incorporating learning-by-researching effect, industry-level scale effect, equipment-level scale effect, the effects of material price, and capital price.

Considering the impact of cumulative knowledge on unit power generation costs, it is possible to extend Equation (14) to Equation (16).

$$\mathrm{l}\mathrm{n}\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}=\alpha \mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+\beta \mathrm{l}\mathrm{n}\mathrm{R}\mathrm{D}+\mathrm{l}\mathrm{n}{n}_0$$

(16)

$$\mathrm{L}\mathrm{B}\mathrm{R}=1-2^{\beta }$$

(17)

where $RD$ is the cumulative R&D expenditure or the number of patents from the first year to the $t$ year. $\beta$ is the elasticity coefficient of learning-by-researching and reflects its learning effect. $\mathrm{L}\mathrm{B}\mathrm{R}$ is the learning rate of learning-by-researching. Cumulative experience and cumulative knowledge generally exhibit increasing trends over time. When these two factors are used as explanatory variables, their high correlation can result in unstable outcomes within the models [33]. Consequently, numerous studies neglect to account for the learning effects of learning-by-researching [7,9,15]. This paper refers to these studies and omits the learning effects of learning-by-researching. The correlation coefficient matrix illustrating the relationships among cumulative installed capacity, cumulative number of patents, and annual power generation for onshore wind and photovoltaics in China is presented in

Table A1. Descriptive statistics on the drivers of LCOE for onshore wind.

Variable	LCOE	CAP	PAT	CPG	Y	K	M
LCOE	1.000
CAP	−0.977 ***	1.000
PAT	−0.974 ***	0.997 ***	1.000
CPG	−0.967 ***	0.987 ***	0.996 ***	1.000
Y	−0.899 ***	0.956 ***	0.969 ***	0.974 ***	1.000
K	0.827 ***	−0.784 ***	−0.763 ***	−0.735 ***	−0.612 **	1.000
M	−0.379	0.379	0.390	0.437	0.470	−0.027	1.000

*** p < 0.01, ** p < 0.05.

and

Table A2. Descriptive statistics on the drivers of LCOE for solar photovoltaics.

Variable	LCOE	CAP	PAT	CPG	Y	K	M
LCOE	1.000
CAP	−0.744 ***	1.000
PAT	−0.786 ***	0.997 ***	1.000
CPG	−0.721 ***	0.999 ***	0.994 ***	1.000
Y	−0.672 **	0.954 ***	0.957 ***	0.954 ***	1.000
K	0.881 ***	−0.711 ***	−0.739 ***	−0.688 ***	−0.605 **	1.000
M	0.847 ***	−0.361	−0.417	−0.334	−0.301	0.658 **	1.000

*** p < 0.01, ** p < 0.05.

of Appendix B. Specifically, the correlation coefficient between the cumulative installed capacity and the cumulative number of patents for both onshore wind power and photovoltaics is 0.997 at the 99% confidence level.

While the learning effect of learning-by-researching can reasonably be omitted from analyses of costs, a considerable number of studies have shown that considering only the effect of cumulative installed capacity on the decrease in LCOE leads to a significantly higher estimated learning rate for learning-by-doing [14,41,42,43]. This paper refers to the work of Wilson [34] and Yu et al. [15] and further incorporates scale effects into its model. The logarithm of Equation (11) is taken to obtain Equation (18).

$$\mathrm{l}\mathrm{n}\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+{\displaystyle \frac{1-r}{r}}\mathrm{l}\mathrm{n}Y+\mathrm{l}\mathrm{n}{n}_0$$

(18)

Using the current year’s power generation (CPG) to characterize scale $Y$, we found that the correlation coefficients between the CPG of onshore wind and photovoltaics and the cumulative installed capacity are found to be 0.987 and 0.999 (99% confidence interval), respectively. Consequently, the cumulative installed capacity ($CAP$) also reflects the CPG.

New power generation projects are rapidly becoming larger in scale. The extant literature also suggests that larger equipment reduces unit power generation costs [54]. Consequently, this paper uses the capacity of a single set of electricity-generating equipment to delineate scale $Y$ in the analysis of scale effects.

Given the above models, we further consider the impact of the price of two input elements: capital (abbreviated as $K$) and material (abbreviated as $M$). Equation (10) is thus rewritten as a MFAM as follows:

$$\mathrm{l}\mathrm{n}\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+{\displaystyle \frac{\gamma }{r}}\mathrm{l}\mathrm{n}K+{\displaystyle \frac{\delta }{r}}\mathrm{l}\mathrm{n}M+{\displaystyle \frac{1-r}{r}}\mathrm{l}\mathrm{n}Y+\mathrm{l}\mathrm{n}{n}_0$$

(19)

where $K$ is the price of capital, $M$ is the price of material, and $\gamma$ and $\delta$ represent the input elasticity coefficient of capital and material prices, respectively.

Given that the five regression coefficients in Equation (18) are equal to $n_1$, $n_2$, $n_3$, $n_4$, $n_5$, we can use the following equations to calculate the learning rate, the price effects of capital and materials, and the scale effects, respectively:

$$\mathrm{L}\mathrm{B}\mathrm{D}=1-2^{\alpha }=1-2^{{\displaystyle \frac{n_1}{\left(1+n_4\right)}}}$$

(20)

$$\mathrm{C}\mathrm{a}\mathrm{p}\mathrm{i}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{e}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}=1-2^{\gamma }=1-2^{{\displaystyle \frac{n_2}{\left(1+n_4\right)}}}$$

(21)

$$\mathrm{M}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{l} \mathrm{e}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}=1-2^{\delta }=1-2^{{\displaystyle \frac{n_3}{\left(1+n_4\right)}}}$$

(22)

$$\mathrm{S}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{e} \mathrm{e}\mathrm{f}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}=1-2^{n_4}$$

(23)

$$r={\displaystyle \frac{1}{1+n_4}}$$

(24)

Furthermore, in reference Söderholm and Sundqvist’s research [14], this paper adds a time trend to the model to test whether the learning effect, scale effect, and price effect are overestimated. Furthermore, we also refer to Benkard [55] and attempt to incorporate the forgetting effect into the model to test the experience depreciation effect. As in Equation (19), the annual cumulative experience (represented by cumulative installed capacity) is further expressed by Equation (25) to test the effect of experience depreciation:

$${\mathrm{C}\mathrm{A}\mathrm{P}}_{\mathrm{t}}=\left(1-\delta \right){\mathrm{C}\mathrm{A}\mathrm{P}}_{\mathrm{t}-1}+∆{\mathrm{C}\mathrm{A}\mathrm{P}}_{\mathrm{t}}$$

(25)

where ${CAP}_t$ is the cumulative installed capacity up to the $t$ year, $\mathrm{\epsilon }$ is the experience depreciation rate, ${CAP}_{t-1}$ is the cumulative installed capacity up to the $t-1$ year, and $∆{CAP}_t$ is the new installed capacity in the $t$ year.

The regression model used in this study is summarized in

Table 1. The regression model.

Model	Regression Equation	Influencing Factor
I (OFAM)	$\mathrm{l}\mathrm{n}\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}=\alpha \mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+\mathrm{l}\mathrm{n}{n}_0$	learning-by-doing
II (TFAM)	$\mathrm{l}\mathrm{n}\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+{\displaystyle \frac{1-r}{r}}\mathrm{l}\mathrm{n}Y+\mathrm{l}\mathrm{n}{n}_0$	learning-by-doing + scale effect at the equipment level
III (MFAM)	$\mathrm{l}\mathrm{n}\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+{\displaystyle \frac{1-r}{r}}\mathrm{l}\mathrm{n}Y+{\displaystyle \frac{\gamma }{r}}\mathrm{l}\mathrm{n}K+{\displaystyle \frac{\delta }{r}}\mathrm{l}\mathrm{n}M+\mathrm{l}\mathrm{n}{n}_0$	learning-by-doing + scale effect at the equipment level + price effect
IV (MFAM)	$\mathrm{l}\mathrm{n}\mathrm{L}\mathrm{C}\mathrm{O}\mathrm{E}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+{\displaystyle \frac{1-r}{r}}\mathrm{l}\mathrm{n}Y+{\displaystyle \frac{\gamma }{r}}\mathrm{l}\mathrm{n}K+{\displaystyle \frac{\delta }{r}}\mathrm{l}\mathrm{n}M+\mathrm{l}\mathrm{n}{n}_0+\mathrm{l}\mathrm{n}T$	learning-by-doing + scale effect at the equipment level + price effect + time trend
V (OFAM)	$\mathrm{l}\mathrm{n}\mathrm{I}\mathrm{n}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{e}\mathrm{d} \mathrm{c}\mathrm{o}\mathrm{s}\mathrm{t}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+\mathrm{l}\mathrm{n}{n}_0$	learning-by-doing
VI (TFAM)	$\mathrm{l}\mathrm{n}\mathrm{I}\mathrm{n}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{e}\mathrm{d} \mathrm{c}\mathrm{o}\mathrm{s}\mathrm{t}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+{\displaystyle \frac{1-r}{r}}\mathrm{l}\mathrm{n}Y+\mathrm{l}\mathrm{n}{n}_0$	learning-by-doing + scale effect at the equipment level
VII (MFAM)	$\mathrm{l}\mathrm{n}\mathrm{I}\mathrm{n}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{e}\mathrm{d} \mathrm{c}\mathrm{o}\mathrm{s}\mathrm{t}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+{\displaystyle \frac{1-r}{r}}\mathrm{l}\mathrm{n}Y+{\displaystyle \frac{\gamma }{r}}\mathrm{l}\mathrm{n}K+{\displaystyle \frac{\delta }{r}}\mathrm{l}\mathrm{n}M+\mathrm{l}\mathrm{n}{n}_0$	learning-by-doing + scale effect at the equipment level + price effect
VIII (MFAM)	$\mathrm{l}\mathrm{n}\mathrm{I}\mathrm{n}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{e}\mathrm{d} \mathrm{c}\mathrm{o}\mathrm{s}\mathrm{t}={\displaystyle \frac{\alpha }{r}}\mathrm{l}\mathrm{n}\mathrm{C}\mathrm{A}\mathrm{P}+{\displaystyle \frac{1-r}{r}}\mathrm{l}\mathrm{n}Y+{\displaystyle \frac{\gamma }{r}}\mathrm{l}\mathrm{n}K+{\displaystyle \frac{\delta }{r}}\mathrm{l}\mathrm{n}M+\mathrm{l}\mathrm{n}{n}_0+\mathrm{l}\mathrm{n}T$	learning-by-doing + scale effect at the equipment level + price effect + time trend

.

3.3. Data Source

This paper analyzes onshore wind power and photovoltaic technologies in China from 2010 to 2022. The LCOE, unit installed costs, cumulative installed capacity, cumulative number of patents, and cumulative electricity generation for onshore wind and photovoltaics are sourced from the International Renewable Energy Agency (IRENA) [56], an intergovernmental organization for renewable energy cooperation that provides the latest data and analysis on renewable energy technologies and investments. Cumulative installed capacity is the annual cumulative value of the new installed capacity reported by IRENA. This data is compiled from publicly available data on power generation projects in various countries and companies, which have been standardized and processed.

This paper uses lending interest rates as an indicator of capital prices. Data are sourced from the World Bank’s statistics on China’s lending interest rates from 1980 to 2024 [57]. We use the steel price composite index to reflect the price fluctuations of the primary materials employed in onshore wind power. Data are obtained from MySteel [58], China’s first commodity data service provider to have obtained certification from the International Organization of Securities Commissions. Lastly, with data sourced from IRENA [8], the price of polycrystalline silicon is used to indicate solar photovoltaic material prices.

The scale effect of onshore wind power and solar photovoltaics at the equipment level is measured by the capacity of wind turbines and photovoltaic modules, as determined through data from the IEA [59], Candela [60], and the CPIA [48]. The technical specifications of the wind turbines used in the Chinese market can be found in the Wind Energy Technology Cooperation Program supported by the IEA. The Candela Component Database contains technical parameters for photovoltaic components that are already on the market, while the CPIA provides statistical data on various links in the industry chain, such as polysilicon, silicon wafers, cells, and modules. Due to the lack of 2020 data for wind turbine capacity, 2019 and 2021 data are linearly interpolated to fill in the lack of 2020 data. For photovoltaic modules, the average capacity value for each year is calculated based on the market share of different battery modules, as the CPIA reported. Since the CPIA was established in 2014, data from 2010 to 2013 are from Candela.

Data for China’s GDP for each year are sourced from the World Bank [61], while the national electricity consumption data are provided by the National Energy Administration [62].

The descriptive statistics for the above data are shown in

Table A3. Statistical description.

Variable	Obs	Mean	SD	Min	Max	Source
Wind turbine namely capacity/MW	13	2.27	0.84	1.467	4.27	IEA [59]
Steel price index	13	141.48	28.68	73.03	176.19	MySteel [58]
Onshore wind power cumulative installed capacity/MW	13	156,882.10	98,990.59	29,475.48	334,980	IRENA [56]
Onshore wind power installed cost/(USD/MW)	13	1190.62	125.23	907	1368	IRENA [8]
Onshore wind power LCOE/(USD/kWh)	13	0.048	0.017	0.022	0.072	IRENA [8]
Wind power cumulative patents	13	46,070.92	31,689.900	2561	109,236	IRENA [66]
Onshore wind power cumulative power generation/GWh	13	1,358,374	1,191,313	9,983,613	3,798,535	IRENA [56]
PV module capacity/W	13	313.26	103.28	218.00	518.30	Candela [60] and CPIA [48]
Polysilicon price/(USD/kg)	13	29.15	21.98	10	79	IRENA [8]
PV cumulative installed capacity/MW	13	126,000	130,000	864	392,000	IRENA [56]
PV installed cost/(USD/MW)	13	1489.86	977.99	552.79	3375.97	IRENA [8]
PV LCOE/(USD/kWh)	13	0.100	0.081	0.030	0.272	IRENA [8]
PV cumulative patents	13	103,428.70	76,604.38	17,264	254,760	IRENA [66]
PV cumulative power generation/GWh	13	225,554.90	256,318.40	518.40	754,558.40	IRENA [56]
Lending rate/%	13	4.98	0.86	4.35	6.56	World Bank [57]
GDP/(USD)	13	1.19 × 1013	2.90 × 1012	7.55 × 1012	1.63 × 1013	World Bank [61]
National electricity consumption/GWh	13	62,520.08	13,850.59	41,923	86,372	National Energy Administration [62]

Note: The price data in the above table are the constant price of the US dollar in 2015.

of Appendix B.

4. Results and Discussion

4.1. Onshore Wind

For onshore wind power technology, the curves fitted using Models IV and VIII more closely approximate the actual values of LCOE and unit installed cost compared to Models I and V (Figure 5 and Figure 6). Their adjusted R² values increased from 0.844/0.738 (Models I/V) to 0.974/0.958 (Models IV/VIII).

We analyze the learning effect of onshore wind power technology using OFAM (Models I and V). Our findings indicate that the learning rates for learning-by-doing are 29.14% and 8.55%, respectively (see

Table 2. Onshore wind power regression results.

Variable	I	II	III	IV	V	VI	VII	VIII
	LCOE	LCOE	LCOE	LCOE	Installed Cost	Installed Cost	Installed Cost	Installed Cost
	LBD	LBD + Scale	LBD + Scale + Price	LBD + Scale + Price + Time	LBD	LBD + Scale	LBD + Scale + Price	LBD + Scale + Price + Time
lnCAP	−0.497 (0.061) ***	−0.126 (0.067) *	−0.120 (0.093)	−0.037 (0.545)	−0.129 (0.022) ***	−0.006 (0.021)	0.017 (0.054)	0.674 (0.191) **
lnY		−0.945 (0.153) ***	−0.776 (0.171) ***	−0.826 (0.371) *		−0.313 (0.069) ***	−0.378 (0.089) ***	−0.775 (0.130) ***
lnK			0.332 (0.234)	0.378 (0.389)			−0.0001 (0.136)	0.366 (0.136) **
lnM			−0.225 (0.098) *	−0.225 (0.105) *			0.061 (0.057)	0.064 (0.037)
Time				−0.054 (0.349)				−0.429 (0.122) **
Constant	2.727 (0.720) ***	−0.909 (0.682)	−0.526 (1.501)	−1.444 (6.132)	8.594 (0.258) ***	7.391 (0.305) ***	6.864 (0.872) ***	−0.402 (2.150)
Adj R2	0.844	0.965	0.977	0.974	0.738	0.906	0.898	0.958
LBD	29.14%	0.48%	1.85%	0.45%	8.55%	0.29%	−0.74%	−11.08%
Scale effect		48.06%	41.60%	45.39%		19.50%	23.05%	41.56%
Capital effect			−5.29%	−4.66%			0.00%	−5.87%
Material effect			3.43%	2.68%			−2.66%	−1.00%
VIF		5.22	8.04			5.22	8.04

*** p < 0.01, ** p < 0.05, * p < 0.1.

). The regression results indicate a significant negative correlation between experience accumulation and both LCOE and unit installed cost, and that experience accumulation plays a substantial role in reducing LCOE. Due to differences in the study periods, the learning rate calculated in this paper, which used 2010 as the base year, is relatively higher. In contrast, Zhang et al. and Dai et al., who used 2006 as their base year, found wind power LCOE learning rates of 10.5% and 6.5%, respectively [5,9].

Models II and VI incorporate the impact of the equipment-level scale effect based on the OFAM framework. We find that both the LCOE and the unit installed cost of onshore wind power technology exhibit significant economies of scale. For Models II and VI, incorporation of the equipment-level scale effect into the OFAM framework result in learning rate reductions of 28.66% and 8.26%, respectively. The scale effect has played a significant role in reducing both LCOE and unit installed costs, indicating that the trend towards larger wind turbines contributes to lowering both LCOE and the average unit installed cost. The scale effect manifests in the economy in two ways: First, larger wind turbines enhance wind energy capture rates and utilization efficiency through taller towers and longer blades, thereby reducing LCOE. Second, increased unit capacity lowers equipment cost per megawatt. The above results suggest that the impact of LBD has been overestimated, and the equipment-level scale effect is the key factor driving cost reductions.

In addition, we incorporated the price effects of capital and materials into the TFAM framework. Our analysis reveals a positive correlation between the price of capital and the LCOE. The regression results align with the anticipated outcomes; higher lending interest rates contribute to the increased financing costs associated with onshore wind power project construction, thereby escalating the LCOE of onshore wind power technology. Wilson et al. found that renewable energy projects are characterized by high upfront capital expenditures, and rising interest rates adversely affect the declining costs of renewable energy technologies [63]. The empirical results of this study similarly demonstrate that an increase in loan interest rates will significantly raise the LCOE for onshore wind power. This implies that policymakers should pay close attention to how changes in the financing environment following global interest rate hikes impact the economic viability of renewable energy projects. The regression results also indicate a positive correlation between steel prices and unit installed costs. In reality, as wind turbines, towers, and other components require a large amount of steel, steel price fluctuation considerably raises unit installed costs. These findings are consistent with the anticipated outcomes; however, steel prices are negatively associated with the LCOE of onshore wind, contrasting prevailing expectations. A likely explanation for this is that the material price effect is secondary to the scale effect, and its negative impact on reducing LCOE is overshadowed by the scale effect. This indicates that the scale effect played a dominant role in the rapid expansion phase of onshore wind power and, to some extent, offset the cost pressures stemming from the rising prices of upstream raw materials.

No significant change in the significance level of different influencing factors for LCOE (Model IV) is found after incorporating technological progress into the model. Moreover, the time trend variable is not significant, indicating that the technological progress of the external industry has no significant impact on the LCOE of onshore wind. Model VI shows that the learning effect, scale effect, and capital price effect on technological progress of external industries all significantly impact unit installed cost. This finding suggests that, compared to LCOE, technological progress in external industries exerts a more significant influence on the installed costs of new or renovated wind power projects.

Models III and VII reveal that the decline in LCOE is attributable to the combined effects of scale effects and material price effects; however, learning effects and capital price effects are found to be non-significant. The decline in unit installed costs is attributable to scale effects, while learning effects, capital price effects, and material price effects do not affect this decline. For onshore wind power technology, the range of scale effect is 23%~42%, and the price effect of materials contributes 3.43% to the reduction in LCOE. Overall, scale effects are the most critical factor in reducing the power generation costs of onshore wind power technology. Thus, future investments in onshore wind power projects should prioritize the use of large wind turbine equipment to amplify the cost advantages brought about by scale effects.

4.2. Solar Photovoltaics

For photovoltaic technology, Models IV and VIII yielded the optimal fitting results (Figure 7 and Figure 8), with adjusted R² values of 0.991 and 0.987, respectively. These values are significantly higher than those of Model I (0.967) and Model V (0.947).

The findings in

Table 3. Solar photovoltaics regression results.

Variable	I	II	III	IV	V	VI	VII	VIII
	LCOE	LCOE	LCOE	LCOE	Installed Cost	Installed Cost	Installed Cost	Installed Cost
	LBD	LBD + Scale	LBD + Scale + Price	LBD + Scale + Price + Time	LBD	LBD + Scale	LBD + Scale + Price	LBD + Scale + Price + Time
lnCAP	−0.393 (0.021) ***	−0.334 (0.028) ***	−0.236 (0.040) ***	−0.402 (0.147) **	−0.322 (0.022) ***	−0.254 (0.027) ***	−0.143 (0.042) ***	−0.358 (0.140) **
lnY		−0.485 (0.184) **	−0.602 (0.155) ***	−0.591 (0.152) ***		−0.564 (0.179) **	−0.762 (0.164) ***	−0.747 (0.153) ***
lnK			1.020 (0.271) ***	0.830 (0.310) **			0.866 (0.287) **	0.620 (0.312) *
lnM			0.033 (0.062)	0.047 (0.062)			0.094 (0.066)	0.113 (0.062) *
Time				0.391 (0.333)				0.507 (0.335)
Constant	1.629 (0.227) ***	3.768 (0.838) ***	1.660 (0.795) *	2.960 (1.350) *	10.572 (0.240) ***	13.058 (0.809) ***	11.332 (0.842) ***	13.017 (1.358) ***
Adj R2	0.967	0.979	0.991	0.991	0.947	0.971	0.985	0.987
LBD	23.85%	11.24%	6.30%	10.50%	20.00%	8.67%	3.87%	9.40%
Scale effect		28.55%	34.12%	33.61%		32.36%	41.03%	40.42%
Capital effect			−32.50%	−25.73%			−26.99%	−18.65%
Material effect			−0.91%	−1.31%			−2.63%	−3.17%
VIF		2.75	6.11			2.75	6.11

*** p < 0.01, ** p < 0.05, * p < 0.1.

demonstrate that for all models, the learning effect of learning-by-doing is statistically significant at the 5% level, indicating that the learning effect has played a crucial role in the cost reduction in solar photovoltaics. The learning rate obtained through OFAM in this study is 23.86%, which is close to the learning rate of 18.72% reported by Dai et al. [5].

Further, we assess the influence of the equipment-level scale effect. Models II and VI suggest that the scale effect played a key role in the decline of LCOE and unit installed costs for photovoltaics, contributing 28.55% and 32.36%, respectively—consistent with expectations. The scale effect manifests in the economy as follows: Increased photovoltaic module capacity delivers higher energy output levels, which can partially offset fixed costs, thereby reducing the unit cost of electricity generation for photovoltaic technology.

After integrating price effects into the TFAM framework. Our findings indicate a substantial correlation between the decline in the LCOE and the unit installed costs of solar photovoltaics and the price effect of capital. This correlation is consistent with expectations. In reality, higher interest rates lead to an increase in the unit installed costs of photovoltaic construction projects. Moreover, the price effect of capital is more pronounced in the decline in the unit costs of solar photovoltaics than those of onshore wind. Research indicates that while fluctuations in upstream material prices directly impact the equipment costs of photovoltaic modules, this effect may be offset by downstream learning effects and scale expansion. Consequently, no significant correlation between input material prices and unit installed costs has been identified [15]. Following the incorporation of the price effect, the learning effect in Models II and VI decreased while the scale effect increased. This suggests that the TFAM model, considering only learning and scale effects, also exhibits estimation bias, underscoring that the price effect cannot be overlooked.

The time trend variable is not significant in all models, and the significance levels of different influencing factors remain constant across these models. Although the significance level remains constant after the inclusion of the time trend variable, the magnitudes of the effects of different influencing factors are altered. For instance, the learning rates for both LCOE and unit installed costs increase by 4.20% and 5.53%, respectively, after incorporating the time trend variable. In addition, the scale effect and the price effect of capital decrease, suggesting that some of the decline in LCOE and unit installed costs may have been incorrectly attributed to the scale effect.

Comparing Models III and VII, we find that the reductions in LCOE and unit installed costs for solar photovoltaics can be attributed to learning effects, scale effects, and the capital price effect; however, the material price effect exerts a relatively limited influence on these cost reductions. For photovoltaic technology, the learning effect range for LBD is approximately 3~7%, the scale effect range is approximately 34~42%, and the capital price effect range is –35% to –26%. This suggests that the expansion of photovoltaic module capacity and the reduction in financing costs are the most significant drivers of cost reduction in the photovoltaic industry.

4.3. Robustness Test

Endogenous problems in the estimation of learning rates often arise [64]. The most prominent endogenous issue in this work is that the increase in installed capacity may lead to a reduction in power generation costs due to the accumulation of experience in learning-by-doing, whereas increases in demand due to decreases in the cost of generation may stimulate increases in installed capacity. Therefore, installed capacity is an endogenous variable. The two-stage least squares (2SLS) method is used to estimate the robustness test and to address endogeneity issues.

In the first stage of 2SLS estimation, this paper employs an approach adapted from the existing literature by using the log of GDP and total electricity consumption as instrumental variables [65]. The second stage uses Equation (19) for estimation. The instrumental variable must be correlated with the endogenous variable, and the instrumental variable should be exogenous to the dependent variable. Both GDP and national electricity consumption fulfill the criteria for instrumental variables. Countries with higher GDPs have greater demands for energy, necessitating substantial installed capacities in their energy systems. Consequently, a positive correlation exists between GDP and installed capacity. Additionally, GDP impacts unit power generation costs through installed capacity, thereby satisfying the exogeneity requirement. Based on a principle similar to GDP, the national electricity consumption also meets the conditions for instrumental variables.

The core explanatory variable of cumulative installed capacity in the fundamental regression Equation (19) was substituted with cumulative power generation (Q) to assess the robustness of the learning effect. There is a mismatch between the two variables [9]. The capacity utilization rate of power generation units is influenced by actual conditions; therefore, a completely linear correlation between installed capacity and electricity generation is impossible. Cumulative electricity generation is indicative of the accumulation of production experience and is employed as an explanatory variable for the learning effect. Consequently, to assess the robustness of the explanatory power of the learning effect, we substitute cumulative installed capacity with cumulative power generation (Q) and conduct OLS and 2SLS estimation analyses (refer to

Table 4. Robustness test results.

Variable	Onshore Wind				Solar Photovoltaics
	OLS-Installed Capacity	2SLS-Installed Capacity	OLS-Power Generation	2SLS-Power Generation	OLS-Installed Capacity	2SLS-Installed Capacity	OLS-Power Generation	2SLS-Power Generation
lnCAP	−0.120 (0.093)	−0.150 (0.077) **			−0.236 (0.040) ***	−0.253 (0.033) ***
lnQ			−0.087 (0.055)	−0.098 (0.043) **			−0.214 (0.032) ***	−0.224 (0.026) ***
lnY	−0.776 (0.171) ***	−0.729 (0.139) ***	−0.765 (0.152) ***	−0.741 (0.120) ***	−0.602 (0.155) ***	−0.551 (0.126) ***	−0.554 (0.145) ***	−0.518 (0.116) ***
lnK	0.332 (0.234)	0.278 (0.189)	0.295 (0.223)	0.265 (0.176)	1.020 (0.271) ***	0.952 (0.218) ***	0.836 (0.261) **	0.782 (0.209) ***
lnM	−0.225 (0.098) *	−0.239 (0.078) ***	−0.213 (0.089) **	−0.218 (0.070) ***	0.033 (0.062)	0.015 (0.050)	0.026 (0.057)	0.014 (0.045)
Adj R2	0.977	0.984	0.979	0.986	0.991	0.994	0.992	0.995
Constant	−0.526 (1.501)	−0.058 (1.231)	−0.748 (1.101)	−0.553 (0.870)	1.660 (0.795) *	1.712 (0.631) ***	1.510 (0.718) *	1.537 (0.567) ***
LBD	1.85%	2.78%	1.41%	1.74%	6.30%	7.57%	6.40%	7.21%
Scale effect	41.60%	39.67%	41.15%	40.17%	34.12%	31.75%	31.89%	30.17%
Capital effect	−5.29%	−5.36%	−4.92%	−4.87%	−32.50%	−34.49%	−29.49%	−29.86%
Material effect	3.43%	4.39%	3.41%	3.84%	−0.91%	−0.47%	−0.81%	−0.47%

*** p < 0.01, ** p < 0.05, * p < 0.1.

for results). These analyses are then compared with the fundamental regression analysis.

The significance of the learning effect, scale effect, and price effect remains robust. Excluding the endogeneity of installed capacity increases the learning rates of onshore wind power and photovoltaic technology by 0.93% and 1.27%, respectively, while the impact of scale effects decreases by 1.93% and 2.37%, respectively. The impact of the price effect of capital increases by 0.07% and 1.99%, respectively, while the price effect of materials increases by 0.96% and 0.44%, respectively. Furthermore, upon substituting cumulative installed capacity with cumulative power generation, the learning rates and the extant influence of each factor remain essentially constant.

4.4. Sensitivity to Experience Depreciation

This section refers to the research of Kobos et al. [13] and sets the experience depreciation rates to several values within the range of 2.5% to 20%. A sensitivity analysis with every 2.5% change was carried out to determine the influence of experience depreciation on various effects.

The results show that changes in the experience depreciation rate have little impact on all factors (see Figure 9). The learning rates of onshore wind and solar photovoltaics remain relatively stable under different experience depreciation rates. Moreover, the price effects of capital and materials do not change significantly, and the scale effects are not obvious.

5. Conclusions

This study aims to address the current lack of understanding regarding the mechanisms behind the dynamic decline in renewable energy generation costs. Building upon learning effects and price effects, it introduces scale effects at the equipment level and capital prices to analyze the joint influence of these factors on LCOE. This paper constructs a unit cost function based on the Cobb–Douglas production function and a cost minimization problem and models it as the combined effect of cumulative experience, average equipment capacity, capital prices, raw material prices, and time trends.

We find that, first, while the LBD effect is overestimated, it still contributes to the decline in LCOE for China’s onshore wind and photovoltaic power generation, accounting for 1.85% and 6.30% of the decline, respectively. Second, the reduction in LCOE for both onshore wind and photovoltaic power in China relies significantly on the scale effect at the equipment level, contributing to 41.60% and 34.12% of the reduction, respectively. Compared to models that consider only learning effects, our model, which incorporates scale effects simultaneously, reduces the learning rates for wind and photovoltaics by 28.66% and 12.61%, respectively. This indicates that scale effects explain part of the LCOE decline and are the most critical drivers of cost reduction. Scale effects remain significantly correlated with the decline in generation costs for both wind and photovoltaics, even after controlling for time-varying technological progress. Finally, the decline in China’s onshore wind and photovoltaic power LCOE is significantly influenced by the capital price effect, while the material price effect has a relatively weaker impact, contributing only 3.43% and –0.91%, respectively.

Therefore, the promotion of learning-by-doing approaches, the advancement of the development of larger-scale power generation facilities, and the full leveraging of financial instruments such as preferential interest rates are crucial pathways for renewable energy technologies to offset their high costs. Governments should support projects that demonstrate the capabilities of new technologies such as offshore wind and solar thermal power generation. These demonstration projects provide individuals with the opportunity to gain experience in the field and optimize costs through efficient implementation and operation. Enterprises in these fields are encouraged to research and develop power generation equipment with larger capacities through special funds, tax incentives, and other policies. Moreover, the industry should formulate technical standards for large-scale equipment that promote the development of such equipment and leverage the positive role of scale effects. Furthermore, green finance’s role in promoting the establishment of green insurance mechanisms for renewable energy projects must be taken advantage of to reduce investment risks and attract social capital. The research and development of technologies for recycling and utilizing wind and photovoltaic power generation equipment should be strengthened to lower material costs and reduce resource waste.

This study has several limitations. As to the learning mechanism, due to the influence of collinearity, we considered only the learning effect generated by experience accumulation, thereby omitting the role of learning-by-researching. In addition, the study focused solely on China and did not engage in comparative research involving other countries. Furthermore, the impact of policy variables such as feed-in tariffs on the decline in renewable energy costs was not considered.