اندوژن سازی تحقیق و توسعه و تجربه بازار در مدل ERIS انرژی ـ سیستم "پایین به بالا"

ERIS, an energy-systems optimization model that endogenizes learning curves, is modified in order to incorporate the effects of R&D investments, an important contributing factor to the technological progress of a given technology. For such purpose a modified version of the standard learning curve formulation is applied, where the investment costs of the technologies depend both on cumulative capacity and the so-called knowledge stock. The knowledge stock is a function of R&D expenditures that takes into account depreciation and lags in the knowledge accumulated through R&D. An endogenous specification of the R&D expenditures per technology allows the model to perform an optimal allocation of R&D funds among competing technologies. The formulation is described, illustrative results presented, some insights are derived, and further research needs are identified.

Research and development (R&D) is one of the basic driving forces of technological progress, contributing to productivity increases and economic growth. Although difficult to measure, the payoffs produced by R&D expenditures are high, both at social and private levels (Griliches, 1995). R&D is also one of the variables that government policies may affect, as private companies are likely to not invest enough in R&D from a public interest perspective, particularly in technologies that are promising only in the long run. In the case of energy systems, R&D constitutes a fundamental factor for the successful introduction of new, more efficient and clean supply and end-use technologies and the achievement of economic, safety, environmental and other goals. Therefore, it is important to study the main mechanisms by which R&D investments contribute to cost and performance improvements of individual technologies and productivity increases of the energy system as a whole. By the same token, it is also interesting to gain insights about the optimal allocation of scarce R&D resources, taking into account that such allocation is influenced by expectations of market opportunities. Thus, it becomes necessary to incorporate those mechanisms into the energy policy decision-support frameworks, e.g., in energy-systems optimization models. However, assessing and quantifying the effects of R&D efforts in energy technology innovation is particularly difficult because of a number of reasons, the broad range of R&D activities relevant to energy issues, the variety of institutions carrying R&D, the difficulties in assessing the (central) role played by industrial R&D and the lack of underlying data, among others (see, e.g., Sagar and Holdren, 2002 for a discussion). Moreover, the role of R&D must be examined within the context of the whole energy innovation system, of which R&D activities are only a part. Demonstration and deployment of energy technologies in the marketplace also play a very important role in their improvement, in particular regarding cost reductions (Grübler, 1998, PCAST, 1999 and IEA, 2000, among others). Technological learning plays an important role in technological change. Learning has many different sources, such as production (learning-by-doing), usage (learning-by-using), R&D efforts (learning-by-searching) and interaction between different social actors (learning-by-interacting), among others (Grübler, 1998). There are a number of technical, social, economical, environmental and organizational factors that influence the presence (or absence) and rate of technological learning processes. The typical representation of this phenomenon is through learning, or experience, curves. The standard learning curve considers the specific investment cost of a given technology as a function of cumulative capacity or cumulative production, which is used as an approximation for the experience accumulated when the technology is deployed. The formulation reflects the fact that some technologies experience declining costs as a result of their increasing adoption (Argote and Epple, 1990). As such, it takes into account the effects of experience due to actual deployment of technologies but it does not provide a mechanism to capture explicitly the effects of public and private R&D efforts, which also constitute an essential component of cost reductions and performance improvements, particularly in the early stages of development of a technology. There is a need to incorporate R&D activities within the technological learning conceptual framework. R&D and market experience can be thought of as two learning mechanisms that act as complementary channels for knowledge and experience accumulation (Goulder and Mathai, 2000). Both mechanisms play an important role. R&D is critical at early stages of development and to respond to market needs, but market experience is essential to achieve competitiveness. There are also feedbacks between these two learning mechanisms. Successful R&D may increase the possibilities of a particular technology to diffuse. Market experience, on the other hand, may contribute to increment the effectiveness of R&D efforts, helping to target them towards needs identified when manufacturing and using the technology. Examples of this interaction have been described in the literature. Neij (1999) and Loiter and Norberg-Bohm (1999), for instance, discuss the case of wind turbines. As a rule, experience gained with deployment of capacity seems to have been critical for progress in wind turbines, having also an influence in the effectiveness of R&D efforts. R&D programs seem to have been more successful when addressing specific problems made evident by the operation experience (Loiter and Norberg-Bohm, 1999). Having a market where new R&D results could be tested was an important feedback mechanism for research and focusing on concrete challenges allowed a more agile and wide incorporation of the innovations produced in such programs in subsequent generations of the technology. Watanabe (1999) performed an analysis of the role of public and private R&D expenses and industrial production in the competitiveness of solar photovoltaics in Japan and, on such basis, they identified the existence of a “virtual cycle” or positive feedback loop between R&D, market growth and price reduction which stimulated its development. Thus, a comprehensive view of technological learning processes and associated policy measures must encompass Research, Development, Demonstration and Deployment (RD3) activities (PCAST, 1999), since all of them play a role in stimulating energy innovation and in the successful diffusion of emerging energy technologies. Energy technology RD3 strategies require, among other actions, a combination of “technology push” and “demand pull” policy measures. On the “technology push” side, well-defined technology roadmaps and strategic R&D portfolios that conciliate short-term and long-term needs may contribute to make technologies available that could enable the provision of energy services in a cleaner, more flexible and reliable way and that can respond to objectives such as climate change mitigation and sustainability. On the “demand pull” side, buy-down policies, procurement and market transformation programs, for instance, could support cleaner and more efficient energy supply and demand technologies, which are currently expensive but with a promising learning potential (Payne et al., 2001, Neij, 2001 and Olerup, 2001). Such policies could contribute to finance the “learning investments” (also called maturation costs), i.e., the investments necessary for these technologies to move along their learning curves until they become competitive. However, R&D productivity is difficult to measure, not least because the observable variables can provide only a partial view of the innovation process. R&D expenditures are used as one of the typical measures of R&D activity. However, there are obstacles in establishing cause/effects relationships between R&D expenditures and technological progress, since R&D expenditures measure an input to the innovation process and not its output(s). In addition, even gathering R&D expenditures can be difficult, particularly for industrial R&D activities. In addition, sound models for the role of R&D in the energy innovation system are not yet available. Clearly, because of the multiple feedbacks between the different factors, a linear model of innovation cannot be established (i.e., with R&D exclusively preceding market experience). However, there is a need for defining, if possible, basic stylized causal rules of interaction between R&D and market experience and their respective effects on technological progress, e.g., cost reductions and/or performance improvements. Regarding the latter, one of the difficulties is that R&D results may not necessarily contribute to the progress of a single technology but to that of several products or services. Different approaches to model the R&D factor as an endogenous driver of technological change in “top-down” and “bottom-up” models have been reported in the literature (see e.g., Grübler and Gritsevskyi 1997, Kouvaritakis et al., 2000a, Kouvaritakis et al., 2000b, Goulder and Mathai, 2000 and Buonanno et al., 2000). In “top-down” models, such as the one presented by Buonanno et al. (2000), the representation is normally through a general knowledge stock that depends on R&D expenditures and is incorporated as a production factor in the production function. Such knowledge stock affects productivity and emission coefficients. In “bottom-up” approaches, the different formulations try to establish a link between these two factors and cost reductions of individual technologies. For such purpose, modifications of the standard learning curve have been proposed. Grübler and Gritsevskyi (1997) present a stochastic optimization micro model, which incorporates uncertain returns on learning due to both R&D and market investments. For that purpose a modified learning curve is used. Such a curve considers cumulative expenditures instead of cumulative capacity as the proxy for accumulation of knowledge. Expenditures in both R&D and commercial capacity deployment are added up to contribute to the cumulative expenditures. Such an approach considers the two factors as complementary and it has the advantage of measuring both factors in common (monetary) units. However, it does not allow for differentiating their contributions. That is, one monetary unit of R&D produces the same effect as one of cumulative market investments. Kouvaritakis et al., 2000a and Kouvaritakis et al., 2000b) have applied the so-called two-factor learning curve (hereon referred to as 2FLC) concept in POLES, a system dynamic, behavioral-oriented model where technological learning is driven by adaptive expectations (i.e., without perfect foresight). The 2FLC is an extension of the standard learning curve, which is based on the hypothesis that cumulative capacity and cumulative R&D expenditures drive the cost reductions of the technology. In such 2FLC formulation, the specific cost of a given technology is a function of cumulative capacity and cumulative R&D expenditures. Such a function is assumed to be of the same kind of a Cobb–Douglas production function, with both factors acting as substitutes according to their corresponding so-called learning-by-doing and learning-by-searching elasticities. The ERIS (Energy Research and Investment Strategy) model was developed as a joint effort between several partners within the EC-TEEM project.1 ERIS is a perfect-foresight energy-systems optimization model. It provides a stylized representation of the global electricity generation system and endogenizes learning, or experience, curves. The original specification was made by Messner (1998) and implemented by Capros et al., 1998 and Kypreos, 1998 and Kypreos and Barreto (1998). A detailed description of the model may be found in Kypreos et al. (2000). Analyses using ERIS have been reported in Barreto and Kypreos (2000). Here, a modified version of the 2FLC, which incorporates the concept of knowledge stock instead of cumulative R&D expenditures, is implemented in ERIS. In doing so, we recognize the limitations posed by the 2FLC hypothesis and the unsolved estimation and data issues associated with it, but emphasize the fact that it constitutes an important step towards the understanding of the role of R&D in energy innovation and its conceptual treatment in energy systems models and the fact that the work has helped to identify a number of research needs in this area. Additional analyses applying the formulation of ERIS with 2FLC developed here are presented by Miketa and Schrattenholzer (2001). The remainder of this paper is structured as follows. First, the standard formulation of learning curves incorporated in ERIS is briefly described in Sect. 2, in order to provide a reference for the developments presented here. Then, the concept of knowledge stock is introduced in Sect. 3. Subsequently, the implementation of the 2FLC in ERIS is presented in Sect. 4. Sections 5 to 7 present and discuss some illustrative examples. Finally, some concluding observations and research needs are outlined in Sect. 8.

A modified two-factor learning curve formulation is implemented in the energy-systems optimization ERIS model and some illustrative modeling results are presented. The formulation allows for considering the effects of R&D together with those of market experience in the learning process of energy technologies. The model finds the optimal allocation of a given R&D budget across a set of competing learning technologies using the two-factor learning curve as the guiding allocation rule. The endogenous specification of R&D expenditures makes the allocation of R&D resources dependent on other parameters and variables of the model, such as carbon constraints, specified market penetration constraints, demand growth, etc. The explicit incorporation of R&D in energy-systems models is important for providing a more comprehensive picture of the technological learning process. Clearly, empirical evidence shows that Research, Development, Demonstration and Deployment (RD3) activities are all important in the energy innovation process and in the successful diffusion of emerging energy technologies. Thus, a more adequate representation of the energy innovation process in the modeling frameworks can be useful, among others, to conduct a more complete examination of energy technology policies. Model analyses may produce insights into how to invest scarce R&D resources more effectively and, thus, they could contribute to more systematic efforts in conforming robust and flexible portfolios of promising new energy supply and demand technologies whose development should be supported. The knowledge accumulated through R&D efforts is represented here by a knowledge stock function. Such function allows for considering retards between R&D spending and productivity gains (in this case cost reductions) and the fact that past R&D investments depreciate and become obsolete. Through the depreciation rate, a “forgetting-by-not-doing” feature is introduced in the R&D component of the learning process. Leaving aside the effects of accumulating capacity, “forgetting-by-not-doing” implies that if no efforts on R&D are made on a given technology its investment costs may increase. In our particular framework, assumptions concerning the rate at which knowledge depreciates alter significantly the dynamics of allocation of R&D funds. Specifically, faster knowledge depreciation may favor allocating more funds to currently competitive technologies in order to avoid or mitigate their “forgetting” process, rather than allocating them to currently expensive technologies that are promising only in the long run. The possibility of introducing such “forgetting-by-not-doing” characteristic also in the cumulative capacity component of the learning process should be examined carefully. The approach depends critically on obtaining a statistically meaningful estimation of separate learning-by-doing and learning-by-searching indexes. Problems regarding the quality of the underlying data and the estimation itself remain to be solved. For instance, multi-colinearity — that is, high correlation between the two explicative variables (i.e., cumulative capacity and cumulative R&D expenditures or knowledge stock) — arises. One of the reasons for such high multicolinearity can be the fact that each of those variables may respond to changes in the other. Increases in the sales volume, for example, may trigger a higher R&D spending by producing firms, so as to ensure the technology remains competitive in the marketplace. On the other hand, R&D breakthroughs may increase the acceptability of the technology in the market or enable the introduction of new products or services. Methodologies to deal with the estimation problems should be developed. Some promising results appear to have been obtained applying panel data analysis for a set of different countries (Klaassen et al., 2002). Possible drawbacks of the 2FLC implementation presented here should be analyzed more carefully and alternative approaches should be explored. In connection with this issue, some authors (Watanabe et al., 2002) have pointed out the possibility that the current formulation of the 2FLC implies “duplication” of the factors. The rationale behind such argument is that the cumulative capacity factor already accounts for some product-embodied knowledge stock. Therefore, considering a separate knowledge stock would drive to some double counting of its effects. In this line of arguments, the cumulative capacity factor should be “corrected” to discount the effects of the product-embodied knowledge stock. This proposition deserves further scrutiny. Alternative approaches have been suggested. Kram (2001) proposes a linear relationship between the learning rate of a single-factor learning curve and a measure of the R&D intensity, which basically assumes that increasing R&D intensity will increase the learning rate of the technology. This relationship has been applied in an exogenous way in the MARKAL model (Fishbone and Abilock, 1981) to assess the impact of additional R&D on the penetration of a given technology. As discussed above, although the use of knowledge stock provides a more complete and sophisticated treatment, its data requirements can be more intensive. If this is so, approaches based on R&D intensity could be favored. However, additional work is required to examine this approach more carefully. On the one hand, the empirical evidence should be analyzed. In particular, it is important to examine whether microstructure changes in the slope of the learning curve (i.e., changes in the learning rate) can be associated with changes in the level of R&D intensity or R&D expenditures. On the other hand, the effects of an endogenous representation of this type of relationship in the models should be examined. Still, with the limitations and unsolved issues, the introduction of this second factor into the learning curve enables an improved and more comprehensive (though, of course, not complete or definitive) treatment of the factors involved in the cost reduction and allows the modeler to take into account the effects of R&D in energy technology policy in a more direct way. In such sense, this work constitutes a first step towards the incorporation of mechanisms that capture the effects of R&D efforts in the technological progress of energy technologies in the ERIS model. Traditionally, such effects have been either ignored or modeled in an exogenous way. For instance, awareness of actual, or consideration of future plans for an increased R&D spending could drive to more optimistic considerations regarding future cost and efficiency trends for a particular technology. Also, when applying the standard single factor learning curve, R&D could be reflected as a factor influencing the starting point of the learning curve or the corresponding learning index (see, e.g., Seebregts et al., 1998). Thus, its explicit incorporation in the learning curve and endogenous formulation in the model provide more “degrees of freedom” as to the way its impact and related policy questions may be addressed. But, increased “degrees of freedom” will very likely imply increased data requirements and, in the absence of reliable data, they will drive to a mounting number of assumptions. Although this certainly will pose difficulties, it should not be a discouraging point. The concept of two-factor learning curves and the work around it have pointed out the need for evaluating the effects of energy R&D investments within the context of technological learning. Moreover, it highlights the need to collect the relevant data, conduct the case studies necessary to evaluate the missing variables and advance in the specification of sound theoretical models. Among other issues, further work should be devoted to a more elaborated representation of the process of allocation of R&D resources. If possible, the contributions of public and private actors should be differentiated. Also, the possibility of introducing stylized considerations concerning the influence of the technology’s life cycle in the relative contributions of market deployment and R&D efforts should be explored. For instance, although in the examples described here both mechanisms act simultaneously for all the learning technologies, one could also consider situations where knowledge can be accumulated on a given technology through R&D before capacity deployment takes place. In addition, although the formulation applied here treats both contributing factors as substitutes, some degree of complementarity between them very likely exists. The examination of their substitutability and/or complementarity characteristics and how they can alter the basic formulation given here is an aspect that deserves a more profound analysis. In addition, the approach followed here is deterministic. However, long-term future technological developments are highly uncertain and the outcome of technological change processes — in particular the emergence of radical innovations and the “winners”, i.e., the technologies that will actually make it to the market — are difficult to predict. Therefore, efforts must be devoted to incorporate uncertainty in the learning characteristics of the different technologies and in other variables in the modeling framework (Grübler and Gritsevskyi, 1997). Another important issue concerns technological learning spillovers across different regions. That is, the fact that a different world region may benefit from the learning efforts of another region on a given technology. The increasing flows of knowledge and technology across world regions and the rising role of transnational energy technology manufacturers and multi-purpose and highly integrated international energy services companies tend to favor the presence of spillovers at the international level. R&D spillovers play an important role (Papaconstantinou et al., 1998) and should be considered and examined carefully. In this area there seems to be a number of important topics to be addressed. Studies at the firm level (Cohen and Levinthal, 1989) have shown that firms perform R&D both to keep their own innovative capacity and to be able to assimilate R&D results from other firms, i.e., to profit from learning spillovers. This argument could be extended to the interactions between different world regions. The effects of international spillovers of energy-related R&D will most likely depend on the assimilative capacity of the different regions (e.g., according to the strength of their own science and technology systems). Attempts should be made to capture this interaction, even if only in a stylized way, in energy-systems models. Other approaches for the incorporation of both learning mechanisms should also be explored. One alternative is the combination of “top-down” and “bottom-up” approaches. As mentioned above, some analyses with “top-down” models (e.g. Buonanno et al., 2000 and Buchner et al., 2002), endogenize technical change in the form of a general R&D knowledge stock that acts as one of the production factors in the economic production function. The R&D knowledge stock enhances the rate of economic productivity and reduces the level of emissions. Technical change can be induced through R&D spillovers across regions. Although this approach provides a way to capture the generic effects of R&D, it does not consider the effects of “learning-by-doing” in specific technologies. The feasibility of combining this type of “top-down” representation of the R&D process with a “bottom-up” model that endogenizes the “learning-by-doing” effect through standard learning curves is worth exploring. It is still early to establish whether the two-factor learning curve will prove a convenient and sound aggregate model adequately supported by the empirical evidence or sound theory. But, even so, it must be understood as a helpful step towards the development of a more consistent representation of the technological learning process, where both market deployment and R&D efforts contribute to the progress of technologies and interact with each other and other model parameters and variables in a common framework. In addition, this work has made more tangible a number of issues that should be tackled by future research efforts. Clearly, there is still a long way to go in disentangling the role of R&D in the energy innovation system. Substantial efforts should be devoted to address the multiple aspects of this problem.