Copyright by Xue Gao 2019
The Dissertation Committee for Xue Gao Certifies that this is the approved version of the following Dissertation:
Essays on Innovation Policy, Knowledge Networks, and Cost Reductions in Deployment-Related Technologies in the Solar PV Industry
Committee:
Varun Rai, Supervisor
Sheila Olmstead
Kenneth Flamm
Gregory Nemet
Essays on Innovation Policy, Knowledge Networks, and Cost Reductions in Deployment-Related Technologies in the Solar PV Industry
by
Xue Gao
Dissertation
Presented to the Faculty of the Graduate School of The University of Texas at Austin in Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosophy
The University of Texas at Austin August 2019
Dedicated to my parents, my husband Fei, and our daughter Adeline. Acknowledgements
I would like to express the deepest appreciation to my committee chair Dr. Varun Rai, who gave me continuous guidance and unlimited support to help me go through all the difficulties I came across throughout my Ph.D. career. He encouraged me to explore big questions and learn cutting-edge methodologies. He guided me on how to write and publish academic papers. He supported me to present our research in front of experts at academic conferences. More importantly, he always had confidence in me and never doubted my abilities. Under his supervision and support, I gradually built my confidence in being a good researcher and teacher. Without his support, I could not have done what I was able to do.
I would also like to extend my sincere gratitude to my committee members, Dr. Sheila Olmstead, Dr. Kenneth Flamm, and Dr. Gregory Nemet, for their valuable advice and unwavering support. It is very fortunate of me to have them as my committee members, who have always been generous with their time, ideas, support, and recommendation letters. I am also grateful to Dr. Joshua Busby, who was always eager to help. He was the most influential person in developing my qualitative skills. I have greatly appreciated his time, support, guidance, and goodwill throughout my job search. I also want to thank Dr. David Eaton for providing me with encouragement and guidance throughout the duration of the title insurance project.
I’m extremely grateful to Dr. Victoria Rodriguez. She is a motherly figure for me and helped me go through many difficulties I had during my Ph.D. journey. I would also
v like to thank the professors at the LBJ school who have taught and helped me: Dr. Peter Ward, Dr. Bill Spelman, Dr. Chandler Stolp, Dr. Paul von Hippel, and Dr. Carolyn Heinrich. They were always there to provide helpful advice, unparalleled support, and academic resources.
Special thanks to outstanding members of the Rai group, Cale Reeves, Ariane Beck, Mark Hand, Erik Funkhouser, Noorah Alhasan, Ashok Sekar, Vivek Shastry, Tiffany Wu, and Sindhu Maiyya. They have been friends, colleagues, and mentors whom I have treasured so much. I have greatly appreciated their professionalism and support as we worked together. The pleasant memories of weekly group lunch and many discussions and chats with them will be hard to forget.
I must thank collogues at the LBJ school, Cale Reeves, Francisca Bogolasky, Jodi Rosenstein, Nisha Krishnan, Miha Vindis, Noah Dust, Ilse Oehler, Abby Lane, Santiago Tellez, Christina Caramanis, Jiameng Zheng, Mark Hand, Rachael Singer, Selena Caldera, Ashlyn Webb, Diana Bolsinger, Eun Young Kim, Alfonso Alvarez, Julian Plough, and Regina Buono. Being part of the LBJ school has given me many pleasant memories to cherish. I want to thank them for their kindnesses and unwavering support.
Finally, to thank the people who shaped me into the person I am. To my parents, I’m extremely grateful that you’ve always inspired me to pursue my dream, encouraged my ambitions, stood behind me, provided continued support, and guided me in the right direction. I salute you for the selfless love and sacrifice you did for me. You mean the world to me, and I will never be able to thank you enough. To my husband Fei, you were always there for me without a doubt. You provided me with endless love, support, and vi companionship throughout the whole process. I am blessed to have you by my side. To my beloved daughter Adeline, I am forever grateful to have you in my life. Thank you, my sweet girl, for joining us when I started writing the dissertation. Thank you for giving me immense happiness and for motivating me to keep reaching for excellence.
vii Abstract
Essays on Innovation Policy, Knowledge Networks, and Cost Reductions in Deployment-Related Technologies in the Solar PV Industry
Xue Gao, Ph.D. The University of Texas at Austin, 2019
Supervisor: Varun Rai
Abstract: Deployment policies have stimulated a large number of renewable energy investment and deployment on a global scale, but they still need to be scaled up at least six times faster to keep global warming below hazardous levels. Therefore, particular emphasis should be placed on understanding the impact of deployment policy on technological innovations and cost reductions, because being more cost-effective and efficient can further facilitate scaling-up. The overarching aim of the dissertation is to understand whether and how deployment policy can lead to technological changes and cost reductions. To be specific, the first essay focuses on whether and how demand-pull policy can impact technological innovation. The second essay explores why deployment policy is essential to promote technological innovations through studying firms’ knowledge acquisition behaviors when conducting innovations. The third essay investigates three primary learning mechanisms through which deployment policy can facilitate cost reductions and contributes to separating these three mechanisms, namely, learning by doing, learning by searching, and learning by interacting. The particular focus of this dissertation is the role of geography. The importance of geography as an element of technological learning and viii innovation continues to be debated by researchers. The salience of this topic is ever greater today, when governments across the world are struggling to balance the opposing pulls of inward-looking, protective approaches to economic growth (e.g., “local jobs” and “preferential procurement”) and outward-looking, collaborative, and open approaches to economic growth (e.g., “open innovation” and “imported talent”). Additionally, this dissertation is related to a broader question, which also sums up my main research topic: what combination of innovation ecosystem features – including, local deployment policies, technology characteristics, social and human capital, demand factors, and supply conditions -- translate into persistent localized innovation and economic benefits.
ix Table of Contents
List of Tables ...... xiv
List of Figures ...... xvii
Chapter 1: Introduction ...... 1
Background ...... 1
Research objectives ...... 4
Empirical focus and research questions ...... 6
Brief summary of each essay ...... 8
Chapter 2: Relevant literature ...... 13
Technological innovation and innovation policy ...... 13
Technological learning ...... 20
Why and how geography matters in innovation ...... 22
Soft costs reductions and technological innovations in solar energy ...... 26
Patent data and citation analysis ...... 29
Chapter 3: Local demand-pull policy and energy innovation: Evidence from the solar photovoltaic market in China ...... 33
Introduction ...... 33
Background and related literature ...... 36
Related literature ...... 36
Background on distributed PV policies in China ...... 39
Data and methodology ...... 42
Data ...... 42
Measurement ...... 45
Empirical models ...... 47 x Omitted variables and potential endogeneity ...... 51
Results ...... 51
Descriptive analysis of PV BOS patents in China ...... 51
Regression model results ...... 55
Discussion and limitations ...... 61
Conclusion and policy implications ...... 63
Chapter 4: Does Geography Matter? A Study of Technological Learning and Innovation in the Solar Photovoltaic Balance-of-Systems (PV BOS) Industry ...... 68
Introduction ...... 68
Theoretical background ...... 70
Data and methodology ...... 74
Data ...... 74
Methodology ...... 76
Gravity model ...... 76
T-tests ...... 78
Regression model ...... 79
Results and discussion ...... 82
Descriptive analysis ...... 82
Results and discussions ...... 87
Distance and border effect on citation networks ...... 87
T-test results ...... 90
The effect of local/nonlocal knowledge on patent quality ...... 92
Conclusion and policy implications ...... 97
xi Chapter 5: Separating multiple learning mechanisms that facilitate cost reductions: Evidence from the U.S. solar PV industry ...... 101
Introduction ...... 101
Theoretical background ...... 104
Learning curve and learning mechanism ...... 104
Supply chain collaboration ...... 106
Data and methodology ...... 108
Data ...... 108
Econometric models ...... 110
Results and discussion ...... 117
Network characteristics ...... 117
Regression results ...... 124
Conclusion and policy implications ...... 132
Chapter 6: Conclusion ...... 137
Contributions to the literature ...... 137
Policy implications ...... 139
Future directions ...... 141
Appendice ...... 143
Appendix A: Supplemental information for Chapter 3 ...... 143
Robustness check ...... 143
Summary statistics ...... 153
Three ways of grouping the 31 provinces in China ...... 155
Specific items for PV BOS classification ...... 157
Appendix B: An analysis of US BOS innovations ...... 158
xii Appendix C: Supplemental information for Chapter 4 ...... 164
Robustness check ...... 164
Summary statistics ...... 169
Appendix D: Supplemental information for Chapter 5 ...... 170
Robustness check: excluding vertically integrated installers ...... 170
Robustness check: Instrumental variables ...... 175
Summary statistics ...... 177
References ...... 178
xiii List of Tables
Table 1: Summary of relevant empirical studies to test whether the location of
demand matters for inducing innovation...... 38 Table 2: Search terms for PV BOS patents in Chinese and English...... 44 Table 3: Regression results for Equations 1 and 2 using different estimation methods. ...56 Table 4: Regression results for different grouping methods (Equation 3), based on competitiveness index according to the city ranking in the “Blue book of
China’s provincial competitiveness,” geography, and GDP...... 57 Table 5: Regression results for differentiation between inverter and non-inverter
patents. Note that Model 7 is the same as Model 2...... 59 Table 6: Regression results for including only inventions in the dependent variable...... 60 Table 7: The specific items and keywords for PV BOS patent data collection ...... 75 Table 8: Comparison between the percentage of knowledge source and percentage of
knowledge diffusion of U.S. PV BOS patents across different countries ....85 Table 9: Comparison between the percentage of knowledge source and percentage of knowledge diffusion of U.S. PV BOS patents across different states in
the U.S...... 86 Table 10: Regression results of the gravity model ...... 89 Table 11: T-test results of the proportion of total backward citations (added by both applicants and examiners) that are in the same country as the focal
patents for different categories ...... 91 Table 12: T-test results of the proportion of total backward citations (added by both applicants and examiners) that are in the same state as the focal patents
for different categories ...... 91
xiv Table 13: T-test results of the proportion of backward citations (only added by applicants) that are in the same country as the focal patents for different
categories ...... 92 Table 14: T-tests results of the proportion of the backward citations (only added by applicants) that are in the same state as the focal patents for different
categories ...... 92 Table 15: Regression results of equation 3 and equation 4 using different windows for
receiving forward citations (exclude self-citations) ...... 95 Table 16: Regression results of equation 5 using different windows for receiving
forward citations (exclude self-citations) ...... 96 Table 17: The specific items and keywords for PV BOS patent data collection ...... 109 Table 18: The annual number of installers, panel manufacturers, and inverter
manufacturers ...... 119 Table 19: The regression results of the effect of learning by doing, learning by
searching, and learning by interacting ...... 128 Table 20: The regression results of including total installed capacity, market share
and competition ...... 131 Table A1: Regression results of including global demand ...... 144 Table A2: Regression results of excluding colinear independent variable ...... 146 Table A3: Regression results of including utility-scale installed capacity ...... 149 Table A4: Regression results of equation 1 and 2 using Poisson regression ...... 150 Table A5: Regression results of equation 3 using Poisson regression ...... 151 Table A6: Regression results for including only inventions in the dependent variable
and differentiating inverters and non-inverters using Poisson regressions .152 Table A7: Summary statistics for key province-level variables...... 153 xv Table A8: Correlation table for independent variables ...... 154 Table A9: Scope of the technologies included within each of the four main BOS
components analyzed in the main chapter...... 157 Table B1: Data sources and descriptions of dependent variables ...... 162 Table B2: Regression results for the equation1 using different year lags ...... 163 Table C1: Regression results of equation 3 and equation 4 using different windows
for receiving forward citations (include self-citations) ...... 165 Table C2: Regression results of equation 5 using different windows for receiving
forward citations (include self-citations) ...... 166 Table C3: Regression results of equation 3 and 4 using Poisson regression ...... 167 Table C4: Regression results of equation 5 using Poisson regression ...... 168 Table C5: Summary statistics for key variables...... 169 Table D1: Installed capacity and the percentage of vertically integrated installers that
use their own inverters ...... 171 Table D2: Installed capacity and the percentage of vertically integrated installers that
use their own panels ...... 172 Table D3: Regression results including the instrumental variable ...... 175 Table D4: Summary statistics for key variables...... 177
xvi List of Figures
Figure 1: Global greenhouse gas emissions by sector from 2000 to 2014 (MtCO2e)...... 2 Figure 2: The solar PV and wind energy capacity (GW) and renewable energy
investment (billion USD) from 2007 to 2017...... 4 Figure 3: The dissertation structure ...... 5 Figure 4: Residential Solar PV System Pricing. Source: SEIA/Wood Mackenzie
Power & Renewables U.S. Solar Market Insight ...... 27
Figure 5: Installed capacity (MW) of distributed PV in China, 2011-2014...... 40 Figure 6: Annual distribution of PV BOS applications, inventions, utility and design
(counts)...... 52 Figure 7: Annual number of PV BOS patents including invention application, invention, utility and design (count), by category and installed capacity
of installed distributed PV (MW)...... 53 Figure 8: Geographic distribution of the provinces with less than or equal to 10 PV BOS patents from 2012 to 2014 (blue) and the provinces with more than
or equal to 60 PV BOS patents from 2012 to 2014 (red)...... 54 Figure 9: Geographic distribution of the provinces with nearly zero MW installed capacity of distributed PV from 2012 to 2014 (blue) and the provinces with more than or equal to 500 MW installed capacity of distributed PV
(red)...... 54 Figure 10: Annual distribution of PV BOS patents by technological category (counts). ..84 Figure 11 Annual distribution of assignee numbers by technological category
(counts)...... 84 Figure 12: The annual number of total installers and the annual number of installers
with PV BOS patents from 2000 to 2014 ...... 118 xvii Figure 13 (a) : An example network structure between two installers and their inverter
manufacturers ...... 120 Figure 13 (b) : An example network structure between two installers and their panel
manufacturers ...... 120 Figure 14 (a) The distribution of installers’ all projects to panel manufacturers and
inverter manufacturers in the market ...... 123 Figure 14 (b) The distribution of installers’ principle projects to panel manufacturers
and inverter manufacturers in the market ...... 124 Figure A1: The geographic distribution of province groups based on geography...... 155 Figure A2: The geographic distribution of province groups based on competitiveness index according to the city ranking in the “Blue book of provincial
competitiveness”...... 155 Figure A3: The geographic distribution of province groups based on GDP...... 156 Figure B1: Annual U.S. PV BOS patents by technology category (count) ...... 159 Figure B2: Annual U.S. PV BOS assignees by technology category (count) ...... 160 Figure B3: The share of U.S. PV BOS patents by state ...... 161 Figure D1: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market. Figure on the left includes all the installers
and figure on the right excludes the vertically integrated installers...... 173 Figure D2: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market (excluding the vertically integrated
installers) ...... 174 Figure D3: The distribution of installers’ projects to panel manufacturers and inverter manufacturers in the market (including all the installers) (the same graph
in the main text) ...... 174 xviii Chapter 1: Introduction
BACKGROUND
The Special Report on Global Warming of 1.5 °C issued by Intergovernmental Panel on Climate Change (IPCC) points out that “human activities are estimated to have caused approximately 1.0°C of global warming above pre-industrial levels in 2017 and global warming is likely to reach 1.5°C between 2030 and 2052 if it continues to increase at the current rate” (IPCC, 2018). Therefore, to limit the global temperature below 1.5 °C by 2030, urgent and ambitious actions on global climate change mitigation are required. The energy sector accounts for more than 70% of the total global greenhouse gas emissions (Figure 1), and thus decarbonizing the energy sector is recognized as a critical way to achieve the long-term climate goals. Phasing out the use of fossil fuel and accelerating renewable energy deployment play a key role in decarbonizing the energy sector. However, achieving a rapid and global-scale deployment of renewable energy requires overcoming various economic, technological, institutional, political, social, and financial barriers. Facing these barriers, governments around the world have adopted various policies to support renewable energy deployment, including but not limited to carbon taxes, cap and trade schemes, feed-in tariff, renewable energy portfolios, and tax credits.
Why is public policy important in helping overcome these barriers and accelerate deployment? While the increasingly urgent climate change challenge may serve as the practical justification for policy intervention in renewable energy deployment, there are several theoretical rationales to justify deployment policy. In particular, policy interventions are needed to correct multiple market failures. The first market failure is environmental externalities. Greenhouse gas emissions have negative externalities, meaning that the social costs of emissions are higher than private costs. Therefore, private investments have insufficient incentives to adopt renewable energy technologies. The second market failure arises because knowledge has positive externalities and thus private benefits of developing new knowledge/technology are lower than
1 social benefits. This also leads to insufficient investment in developing new knowledge and technologies associated with renewable energy. The third market failure is related to adoption externalities, which means that the private costs of adopting new technologies are positively associated with the scale of technology adoptions through mechanisms such as learning by doing and learning by using (Jaffe et al., 2004).
60000
50000
40000
30000
20000
10000
0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
Energy Agriculture Industrial Processes Land-Use Change and Forestry Waste Bunker Fuels
Figure 1: Global greenhouse gas emissions by sector from 2000 to 2014 (MtCO2e). Source: the CAIT Climate Data Explorer. 2017. Country Greenhouse Gas Emissions. Washington, DC: World Resources Institute.
Globally enforced and leakage-free emission taxes and trading system, while widely considered as the most efficient policy instruments to address environmental externalities, face great political barriers and design challenges. Consequently, in a second-best world, deployment policy plays a key role in correcting these market failures and directly increasing technology adoption and diffusion. Deployment policies or demand-pull policies are policy instruments that aim to stimulate investment and market demand. Two common deployment policy instruments for renewable energy technologies are feed-in tariffs (FITs) and renewable portfolio standard (RPS). The FIT is the most common policy instrument, which compensates power generators a fixed-rate
2 based on electricity generation. RPS requires power producers to generate a certain minimum share of their electricity from renewable energy. These deployment policies have largely stimulated renewable energy investment and deployment around the world (IEA, 2018).
In the last decade, the cumulative investment in renewable power and fuels has reached $2.7 trillion (Figure 2), which approximately equals three times the investment in fossil fuel generation over the same time (REN21, 2018). The investment mainly focuses on solar photovoltaics (PV) and wind energy, which accounts for more than 60% of the total investment. Between 2007 and 2017, the global capacity of solar PV technology and wind energy rapidly increased from 8 GW to 402 GW and from 94 GW to 539 GW, respectively. China and the U.S. are the two largest wind and solar PV markets in the world. China accounts for more than half of global demand and manufacturing of solar PV technologies and about 40% of global demand of wind energy. The US placed the second in solar and wind installed capacity. Solar PV became both China’s and the U.S.’s leading source of new generating capacity in 2016 and 2017 (REN21, 2018).
3 600 350
500 300
250 400 200 300 150 200 100 Installed capacity (GW) capacity Installed
100 50 New investment (billion USD)
0 0 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017
Previous year' wind capacity (GW) wind annual additions (GW) Previous year' solar PV capacity (GW) solar PV annual additions (GW) Annual new investment (billion USD)
Figure 2: The solar PV and wind energy capacity (GW) and renewable energy investment (billion USD) from 2007 to 2017. Source: renewables 2018 global status report from REN21
RESEARCH OBJECTIVES
Although deployment policies have stimulated substantial global renewable energy investment and deployment, the renewable energy investment and deployment are still far behind the levels needed to achieve the greenhouse gas emissions mitigation targets over the next 10-15 years (IRENA, 2014, IEA 2018). Therefore, particular emphasis should be placed on understanding how deployment policies can best promote technological innovations and reduce installation costs, making renewable energy technologies more cost-effective, thereby spurring further investments and deployment. It is clear that deployment policies enable scale-up of renewable energy installations, but their impact on technology innovations and cost reductions are much less clear.
4 Therefore, the overarching aim of my dissertation – a collection of three closely-coupled essays – is to understand whether and how deployment policy can lead to technological changes and cost reductions. Specifically, the first essay focuses on whether and how demand-pull policy can impact technological innovation; the second essay, through studying firms’ knowledge acquisition behaviors in the innovation process, explores why deployment policy is essential to promote technological innovations; and, the third essay investigates three major mechanisms through which deployment policy can facilitate cost reductions and contributes to separating these three mechanisms including learning by doing, learning by searching, and learning by interacting (Figure 3). Throughout these three essays, one important focus of my dissertation is the role of geography, the importance of which as an element of technological learning and innovation continues to be debated by researchers. Thus, the overarching focus of my dissertation may be summed up as addressing the following broad question: what combination of innovation ecosystem features – including, local deployment policies, technology characteristics, social and human capital, demand factors, and supply conditions – translate into persistent localized innovation and economic benefits? Essay 2 Essay 1 Demand-pull Technological Policy Innovations Essay 3 Cost Reductions
Supply chain network
Figure 3: The dissertation structure
5 EMPIRICAL FOCUS AND RESEARCH QUESTIONS
The empirical focus of my dissertation is balance-of-system (BOS) components in the solar PV industry. PV has been the fastest growing electricity generation technology globally in the last decade. Large-scale deployment of distributed solar PV technologies is widely considered to be an important piece in addressing the environmental impacts of the electricity sector, making PV an important technology to study. Solar panel/module and non-module components are two major types of hardware in a solar PV system. Solar panels directly convert sunlight into electricity through the photovoltaic effect. BOS components are the major non-module components, which include four technology categories: inverter, mounting, monitoring, and site assessment. Inverters convert power generated by solar PV cells and modules from direct current (DC) to alternating current (AC). Mounting is used to physically support solar PV modules and/or arrays. Monitoring includes methods and devices that monitor health, connectivity, and output of solar PV systems. Site-assessment includes methods and mechanisms for appraising the photovoltaic output potential of a land parcel, determining the ideal placement of PV systems on potential sites, and the extent to which obstructions or shade may impede the output potential of PV sites.
Most academic studies so far have focused on PV panels and modules, and only a small body of literature focuses on BOS. More recently, scholars have begun to emphasize that a wider deployment of solar PV in the future would depend heavily on innovations and cost reductions in
BOS (Ardani et al., 2013, 2012; Barbose and Darghouth, 2018; Seel et al., 2014, 2013). The price of solar panel/module has decreased dramatically over the last 15 years, falling from $3.8 per watt in 2000 to $0.8 per watt in 2015. Currently, panels account for only about 15% of the total price of an installed (distributed) solar PV system. However, BOS component prices haven’t changed substantially. Just by themselves BOS components account for about 20% of the total price of an installed PV system. However, BOS components are also closely related to several non-hardware costs such as installation costs, racking costs, and other installer soft costs (customer acquisition, financing, overhead, etc.). These non-hardware costs account for more than 60% of the total system
6 price. As such, there is a general consensus that the biggest barrier for a wider deployment of PV in the future is the high price of non-module components.
But why is it necessary to study BOS components separately? Firstly, the learning/innovation process for PV panels is different from the learning process of BOS components – a difference mainly driven by different technological characteristics (Lewis and Wiser, 2007; Neij et al., 2017; Shum and Watanabe, 2008; Wene, 2000). PV panels are pretty standard and globally traded products, but BOS components often have much more customized innovations that depend upon the local context such as local climate conditions, local building code, local rooftop structures, local customer’s preference, and local conditions of power grid. Thus, BOS components represent a good case involving the positive, interactive feedback between geography, innovation, and localized economic effects. Technological and operational aspects of BOS components make local conditions salient, thus affording a rich empirical setting to explore the geography-innovation-economy interactions at the local level.
Furthermore, quite interestingly, differences in learning/technology processes are expected to exist even among subcategories of PV BOS components (inverter, monitoring, mounting, and site assessment). Inverter and monitor technologies are expected to have the least local nature, compared to the other three subcategories, because inverters and monitors are closer to globally- traded products that are relatively easily adaptable to the local context. Mounting equipment are expected to have the most local dependence, because mounting techniques, methodologies, and equipment heavily depend on local climate, local building code, and local rooftop features. Site assessment systems and tools are also expected to depend on local conditions, such as local grid features and local customers’ preferences. As a result, different technology characteristics within the four BOS categories allow me to explore how the impact of deployment policy may depend on different technological characteristics and associated learning/innovation processes.
7 In sum, speaking to these theoretical, empirical, and policy opportunities, my dissertation entitled “Essays on Innovation Policy, Knowledge Networks, and Cost Reductions in Deployment- Related Technologies in the Solar PV Industry” explores three major research questions: (1) the efficacy of local demand-pull policy in driving local innovation of PV balance-of-system (PV BOS) components, (2) the knowledge acquisition process that fuels firm-level innovation process in PV BOS, with a focus on linking knowledge networks and innovation quality at different geographic scales, and 3) the role of technological changes and supply-chain networks in cost reductions, with a focus on separating three mechanisms (learning by doing, learning by searching, and learning by interacting) through which deployment policies potentially facilitate cost reductions.
BRIEF SUMMARY OF EACH ESSAY
Essay one ---- My first essay empirically examines whether demand-pull policy can induce technology innovation and whether the locus of demand-pull policy matters for the engendered innovation. I used keyword search strings to develop two original databases of PV BOS patents in the field of distributed PV from the State Intellectual Property Office of The People’s Republic of China (SIPO) between 2005 and 2014 and from the United States Patent and Trademark Office (USPTO) between 2000 and 2014, respectively. The empirical design exploits the subnational variation in (policy-created) demand to study the impact of geographically-differential (local vs. non-local) demand on local innovation. Results from a range of econometric models support the demand-induced innovation hypothesis. More importantly and novel, the results also show that only local demand significantly induces PV BOS innovations in the distributed PV market. These results are consistent in both the U.S. and Chinese markets. At the same time, these results are contrary to what is known for PV modules. My novel focus on PV non-module components, thus, helps complete the picture of how to understand the policy-geography-innovation nexus in PV.
8 Moreover, I explored whether local demand has an indispensable role in inducing local innovation by differentiating between inverter and non-inverter technologies. The results emphasize that within the same technology (an integrated PV system, here) the technological trajectories of different subcomponents (modules, inverters, etc.) potentially interact differently with local market conditions. Local demand has a stronger positive effect on technologies that have a significant local nature. But for relatively standard and globally traded products (such as panels and inverters), technological innovations are more likely to be driven by global markets. These findings suggest that whether local demand is an essential precondition to local industry development (i.e., innovations) depends on technological characteristics and the nature of associated learning processes.
Overall, results from my research emphasize the importance of local demand and show that depending upon the technology involved, local demand-pull policies could translate into local innovative activity. The technological learning fueling innovation in PV BOS results from an interactive process between technology and the local context. In this case, market size is akin to a locally bounded asset, and thus innovation induced by the market is a region-specific competitive advantage. In other words, local demand-side policy can help integrate multiple policy goals including environmental protection, technological innovation, and economic competitiveness. The findings also inform the discussion on ‘green industry development’. The results show that technology characteristics are critical factors in explaining the role of local demand in green industry development, and thus if learning and innovation processes depend on local context, then local demand is a prerequisite for developing a competitive local industry. Additionally, the success of subnational governments’ efforts in expanding market demand and translating it into local innovative activities also underscores the potential viability of a bottom-up approach to renewable energy governance, at least for those technologies that have strong local dependence.
9 Essay two ---- Inspired by my first essay, my second essay explores why innovation is localized and the extent to which localized learning is important for innovation. I answer these questions from the perspective of the underlying learning process with a focus on the characteristics of knowledge acquisition and knowledge networks of PV BOS innovations using patent citation data. For this work, I extended my original U.S. PV BOS database by including backward citations and forward citations for each U.S. PV BOS patent. Backward citations refer to patents that are cited by the focal patent and forward citations include all the patents that cite the focal patent in the future. Using three different quantitative approaches to triangulate the problem, one main contribution of this essay is to quantify the influence of geographic distance on knowledge acquisition, which provides a concrete understanding regarding the extent to which the localized learning is happening in the U.S. PV BOS industry and its importance in solar PV deployment. The results show that citation-related linkages have a strong localization effect at the state level, but not at the country level. Moreover, this study provides empirical evidence showing that knowledge acquisition is associated with technology characteristics. I find that technologies that are more related to local context use more local knowledge, which indicates that technologies with a more significant local nature have a higher dependence on localized learning.
Another important contribution of this essay is to add the international dimension to localization learning literature. I argue that geographic proximity is unlikely to be the whole story of promoting innovations and I indeed find that international knowledge significantly contributes to improving patents’ quality, which emphasizes the value of knowledge diversity in technological innovation. Compared to acquiring local knowledge and building local networks, acquiring non- local knowledge/networks requires extra efforts, costs, and more complicated processes. Further, the difficulties in knowledge acquisition are expected to be more serious for start-ups, firms with low absorptive capacity, and firms located in peripheral regions. Thus, for policymakers, it might be valuable to examine whether local firms face difficulties in accessing non-local knowledge and, if so, to help local firms bridge non-local networks with local ones. For example, depending upon
10 the technological characteristics, it would be useful for local governments or industry associations to organize more formal or informal activities that facilitate international collaboration and knowledge exchange. In sum, the second essay incorporates both local and non-local aspects of technological learning/innovation process, thus providing for a more integrated understanding of the trade-offs between tailoring approaches for cultivating local markets with local strategies and leveraging global learning networks.
Essay three ---- Although technological innovation is intrinsically important, it is often of more direct interest to understand its impact on cost reductions for technologies and services offered on the market. Previous literature on solar costs has focused on how market structure, installer scale, demographics, ownership, policy, and system components influence solar PV installation prices (Gillingham et al., 2016b; Nemet et al., 2016). However, the unexplained variation is still large (>70%) after controlling for these factors, which suggests the need for further work to identify the potential for cost reductions in solar PV installations – a key requirement for much broader deployment of PV at the global scale. To address this research opportunity, my third essay focuses on the role of technological innovation, knowledge spillovers, and upstream- downstream networks (i.e., installer-inverter/panel manufacturer networks) in reducing costs of solar PV installations. My analytical strategy is to match the PV BOS patent database with an extraordinarily rich dataset of system-level PV installations in the U.S., allowing me to build models to explore how innovation and network factors impact installation prices, as well as the embeddedness of innovation and networks in geography.
The second contribution of this essay is to separate three major learning mechanisms through which deployment policy is expected to impact cost reductions. The direct impact of deployment policy is to increase firms’ installation experience. When an installer increases its installation experience, the installer can learn from its repeated working on technologically similar installations to improve its productivity and thus reduce its installation costs (learning by doing).
11 Increasing installation experience can also motivate new working processes, installation methods, or related technology innovations and stimulate firms to invest more in research and development activities, thereby potentially leading to cost reductions (learning by searching). Installers can also learn from its upstream suppliers (i.e., panel manufacturers and inverter manufacturers) regarding new products, strategies, technology trends, and processes, which could also lead to cost reductions (learning by interacting). By incorporating experience-related variables, innovation-related variables, and network-related variables, this essay carefully separates these three mechanisms that convert experience accumulation to cost reductions.
12 Chapter 2: Relevant literature
TECHNOLOGICAL INNOVATION AND INNOVATION POLICY
In 1912, Schumpeter published “the theory of economic development” and put forward “innovation theory,” In Schumpeter’s work, he defined innovation as five types of activities: 1) launch of a new product or a new species of already known product, 2) application of new methods or process of production, 3) acquiring of new sources of supply of raw material or semi-finished goods, 4) opening of a new market, and 5) new industry structure and new forms of organization. As for technological innovation, it mainly refers to the first three activities (i.e., new product, new methods/process, and new materials). These activities can be further distinguished between innovation activities and diffusion activities. Here I mainly focus on innovation activities, and thus it can be defined as “the discontinuity in the technical characteristics or in the use of a new product or new material or process or method” and it includes both radical innovation and incremental innovation (Lundvall, 2005). Schumpeter also proposed that innovation is an important factor in economic growth in 1912. In the following decades, scholars began to contribute to the theory of technology innovation. These neo-Schumpeterian economists including Freeman, Rosenberg, and Nelson emphasized the role of innovation in economic growth. As economists began to more widely recognize that technology innovation is an engine of economic growth, interest in studying the drivers of technology innovation grew.
Hicks was the first to propose the hypothesis that innovation could be induced by “a change in the relative prices of the factors of production”, for example, rising wages can induce labor- saving innovation (Hicks, 1932). Applied to the field of energy, the rising prices of energy could induce energy-saving innovations. Inspired by this idea, environmental economists investigated the effect of rising energy prices or environmental regulations on environment-friendly technology innovation. Many empirical studies demonstrate that rising energy prices or external environmental regulations could significantly stimulate technology innovation (Altwies and
13 Nemet, 2013; Hayami and Ruttan, 1985; Jaffe and Palmer, 2009; Lanjouw and Mody, 1996; Newell et al., 1999; Parry et al., 2003; Popp et al., 2011a, 2010; Taylor, 2008).
Another branch of induced innovation hypothesis, proposed by Schmookler and Griliches, emphasizes that the power of induced innovation lies in market demand (Griliches, 1957; Schmookler, 1962). By studying the technology spillover of hybrid corn, Griliches found that technology spillover responds to market size (Griliches, 1957). Schmookler further concluded that demand not only pulls technology spillover, but also pulls technological innovation. Utterback further pointed out that market factors are the primary influence on innovation, and 60% to 80% of important innovations have been in response to demand (Utterback, 1974). Many recent empirical studies also provide strong evidence to support the demand-induced innovation hypothesis that innovation could be induced by demand (Jaffe and Trajtenberg, 2002; Lanjouw and Mody, 1996; Lin, 2010; Gregory F Nemet, 2009; Popp, 2006a).
The basic model of demand-pull innovation is a linear model that indicates firms’ innovation is triggered by market demand. Roy Rothwell (1983), Fusfeld and Kahlish (1985) and von Hippel (1986) were not satisfied with the linear innovation model, so they examined the innovation process as a process of interactions between different elements, such as between firms, between users and firms, and between government and firms (Fusfeld and Haklisch, 1985; Rothwell, 1983; von Hippel, 1986). Later in 1986, Kline and Rosenberg (1986) proposed the chain-linked model, which is similar to the interaction model (Kline and Rosenberg, 1986). Contrary to the linear model of innovation, the chain-link model studies the complex relationship between different stages in the process of innovation. Each stage has effects on other stages and there are complex feedback loops among all the stages. In this dissertation, I also emphasize and incorporate the interaction view in studying innovation process.
14 Inspired by the interaction idea, scholars began to study technology innovation from a systems perspective, including studies in the “national innovation system” framework. In the 1980s, economies of the major Western industrial countries declined drastically, while the economies of Japan and several emerging countries in East Asia prospered, which was an exciting topic for researchers of technological innovation. Freeman, interested in factors that contributed to the economic growth in Japan, studied the role of governments (especially the Ministry of international trade and industry) and the Japanese strategy of “building the country by technology” in Japan’s technological innovation mechanism, pointing out that the national innovation system contributed significantly to Japan’s high-speed development. Since then, national innovation system has become an increasingly important conceptual framework. The national innovation system framework emphases that innovation system includes various interdependent elements, including firms, industries, governments, universities, complementary infrastructure, non- government organizations, other institutions, and their environments. This framework successfully involves all kinds of institutions, their behaviors, and more importantly, their interactions into innovation analysis and contributes a new rationale to identify technological gaps and economic gaps within an industry, within a country, and across different countries (Godin, 2002; Lundvall, 1992; Lundvall et al., 2002; Motohashi, 2005; Mowery and Sampat, 2006; Mowery et al., 1993; Mowery and Rosenberg, 1993; OECD, 2002).
From the perspective of the micro drivers of innovation, the extant literature provided insightful thoughts of understanding firms’ innovation behaviors, such as why firms innovate and how firms innovate. In order to develop competitive advantages and ensure future competitiveness, improve absorptive capacity (absorb external information/knowledge/technology), increase productivity, reduce costs, and increase profits, firms often invest in innovative activities (Baldwin et al., 2000; Baldwin and Johnson, 1999; Cohen and Levinthal, 1990; Crepon Bruno et al., 1998; Cuervo-Cazurra and Annique Un, 2010; Helfat, 2000; Schumpeter, 1942; Teece et al., 1997). Firms innovate through various activities, including formal investment in research and
15 development (R&D), incremental changes on products and process, imitative activities, and exploiting existing knowledge/technologies in new ways (Evangelista et al., 2002; Grimpe and Sofka, 2009; Hansen, 1997; Huang et al., 2010; Kim and Nelson, 2000; Nascia and Perani, 2002; Veugelers and Cassiman, 1999). Although firms have incentives to innovate, why is it still important to involve government and public policy in the innovation process? Technology policy/innovation policy is a set of public policies that aims at influencing or changing speed, direction, and scale of technological innovation. Next, I summarize and discuss three primary reasons to justify policy interventions on technological innovation.
First, technological innovation policies are necessary to correct market failures. Knowledge is a public good, which has positive externalities. Since firms cannot gain all the benefits from their R&D activities and cannot get paid for knowledge spillover, the firms’ R&D investment will be lower than socially efficient levels of R&D investment. Such conclusions have been agreed upon by many researchers (Arrow, 1962a; David et al., 2000; Griliches, 2013; Nelson, 1959). Another important type of market failure is associated with environmental externalities. Knowledge externality discussed above is a type of positive externality, but environmental externality is a type of negative externality. The private benefits of developing or adopting environmental-friendly technologies to reduce environmental externality are very likely to be lower than social benefits. Therefore, without appropriate policy interventions, firms would not have sufficient incentive to reduce emissions of local or global pollutants. In terms of policy instrument, to put a price on pollutants or emissions through tax or emission trading system is widely recognized as the most efficient policy (Baumol, 1972; Pigou, 1920; Coase, 2013; Stavins, 1995). In a second-best world, government R&D investments and demand-side support for technology innovation are economically justified on the basis of externalities, while it is also recognized that different policy instruments have varying effects on different technologies and those effects may even vary by country (Dechezleprêtre and Glachant, 2014; Hassett and Metcalf, 1995; Molas-Gallart and Davies, 2006; National Research Council, 2001; Peters et al., 2012).
16 Technology/innovation policy may also be more political feasible compared to tax or trading system, because the benefits are typically more concentrated, but the costs are dispersed (Jaffe et al., 2004)1.
Second, government action is needed to correct system failure. As discussed above, innovation is widely recognized as a process of interactions among different elements and the innovation system approach emphasizes that the success of innovation depends on various complementary elements in the system, including conditions for generating and diffusing knowledge, financial conditions, demand for innovation, innovation capacity, and the degree of cooperation (Braun, 2008; Edler and Georghiou, 2007; Fagerberg, 2017; Kline and Rosenberg, 1986; Lundvall, 1992, 1988; OECD, 2002; Smits et al., 2013). The insufficient provision of these complementary elements is recognized as a system failure and the existence of these system failures would block technological changes, and thus policy interventions are needed to identify and correct system failures (Braun, 2008; Edquist, 2011; Fagerberg, 2017; Hoppmann et al., 2014; Klein Woolthuis et al., 2005; Negro et al., 2012; Wieczorek and Hekkert, 2012). Third, even in the absence of the market failures from R&D spillovers or externalities, governments may still want to subsidize innovation in areas that support the national interest (e.g., defense, climate change mitigation and adaptation, security).
There are mainly three types of innovation policy: 1) supply-side policy includes providing technological support and financial support; 2) demand-side policy mainly refers to expanding market size, including government procurement; 3) environmental-side policy includes patent policy, environmental regulation, worker health regulation, and so on (Rothwell, 1982). Although
1 It is worthy to note that it is possible to have socially excessive, inefficient investment in research and development. The “patent races” literature shows that non-colluding firms invest excessively in research and development activities based on the assumption that the winners can capture the largest or all the price and the others receive nothing (Reinganum, 1989; Doraszelski, 2008) . From the perspective of competition and leadership, literature also shows that the incentive of a monopolist to invest in new technology is to preempt entry by competitors, so a monopolist may also invest excessively to be ahead in the competition (Grossman and Shapiro, 1985).
17 public policy is important and necessary in technological learning and innovation, complicated innovation processes limit the effectiveness of public policy tools, thus highlighting the importance of studying technological learning processes that can provide direct insights for improving public policies (Rothwell and Zegveld, 1988). Rothwell and Zegveld (1988) also noted that innovation process is geographically and temporally dynamic. Innovation process can be different in different countries, in different industries and in different sectors of an industry. Therefore, public policies should be flexible enough to adapt to different situations.
As noted above, the main focus of my dissertation is demand-side policy (i.e., deployment policy). Accordingly, below I summarize the policy justification for and the insights from key theoretical and empirical research on the impacts of demand-side policy on technological change. In addition to two types of market failures I discussed in the previous paragraphs, Jaffe et al. (2004) pointed out a third market failure: adoption externalities. Adoption externalities refer to the possibility that private benefits of adopting a new technology might depend on the overall scale of technology adoption (Jaffe et al., 2004). Particularly, the positive benefits of demand-side policy may be acquired from various mechanisms including learning by doing, learning by using, learning by interacting, and learning by searching. These learning mechanisms induced by market demand significantly contribute to cost reductions of technologies. I will provide an in-depth discussion regarding these learning mechanisms in section 2.2.
Many studies focus on demand-pull policies specifically. Learning and innovation often result from an interactive process, which benefits from proximity, thus experience-based innovations are often characterized by localization and geographical agglomeration. The importance of local contexts, such as local culture, local regulation, local business environment, and local supply chain, in the innovation process has been recognized. Local firms have comparative advantages resulting from a better understanding of local context. The comparative advantages of local firms are stronger for the business-to-consumer sectors, as they focus more on
18 end-users and demand-induced innovation (Brem and Voigt, 2009). Demand-pull policy can expand local markets and help local firms gain more experience. However, policy-makers are often more interested in seeing that public support for expanding local markets contributes to local technological innovation and economic competitiveness, rather than benefiting other jurisdictions more heavily (Lewis and Wiser, 2007; Maguire et al., 2012).
Recent empirical work has begun to distinguish the effect of demand-pull policy on local technology innovation and foreign technology innovation, but the results of empirical studies are not consistent. Popp (2006) found that innovations for sulfur dioxide and nitrogen oxide control at coal-fired power plants primarily responds to domestic regulation instead of foreign regulation (Popp, 2006b). Dechezleprêtre and Glachant (2014) analyzed that wind technological improvements across OECD countries respond to both domestic and foreign demand-pull policies, they also suggest that the marginal effect of domestic policies is twelve times greater than foreign policies (Dechezleprêtre and Glachant, 2014). Peters et al. (2012) found that both domestic demand-pull policies and foreign demand-pull policies could spur innovations in solar PV panels, but the magnitude of the effect of foreign demand-pull policies is smaller than that of domestic policies (Peters et al., 2012). On the contrary, Lanjouw and Mody (1996) examined vehicle emission regulations in the U.S. and found that the U.S.’s vehicle emission regulations stimulated innovations in Japan and Germany (Lanjouw and Mody, 1996). Popp et al. (2011) observed that technology innovation in the pulp and paper industry responds to both domestic and foreign regulation (Popp et al., 2011). A possible explanation of these apparently inconsistent results could be whether an industry faces a global market (Popp, 2006b). If an industry’s products focus on foreign markets, it is reasonable to expect that technology innovation in this industry responds to foreign regulation. Another possible explanation is whether the local context is significantly important to technology innovation. PV BOS technologies, the focus of my research, need to interface with and respond to various local contextual factors such as climate, standards and codes,
19 supply chain, and customer preferences. Therefore, based on the existing literature, I expect that the innovations in PV BOS should primarily respond to local demand-pull policies.
TECHNOLOGICAL LEARNING
Technological learning can occur at both the individual level and the firm level. There is a long list of different definitions for technological learning: Carayannis and Kassicieh (1997) defines technological learning as “the process by which a technology-driven firm creates, renews, and upgrades its latent and enacted capabilities based on its stock of explicit and tacit resources” (Carayannis and Kassicieh, 1997); Dodgson (1991) defines technological learning as “the ways firms build and supplement their knowledge-bases about technologies, products and processes, and develop and improve the use of the broad skills of their workforces” (Dodgson, 1991); and Figueiredo (2002) defines technological learning from the perspective of capability, and says “the term technological ‘learning’ is usually understood in two alternative senses. The first sense refers to the trajectory or path along which the accumulation of technological capability proceeds; the second sense refers to the various processes by which knowledge is acquired by individuals and converted into the organizational level” (Figueiredo, 2002). Some other definitions emphasize the interactive nature of technological learning. For example, Bathelt et al. (2004) noted that “Innovation, knowledge creation and learning are all best understood if seen as the result of interactive processes where actors possessing different types of knowledge and competencies come together and exchange information with the aim to solve some – technical, organizational, commercial or intellectual – problems” (Bathelt et al., 2004).
Most learning is not a spontaneous process, but rather a conscious, costly and targeted process (Malerba, 1992). At the firm level, learning can occur at almost every step, including R&D, designing, manufacturing, production, marketing, engineering and deployment (Malerba, 1992). Knowledge and innovation are the most important outcomes of firms’ efforts in learning (Fiol, 1996; Hitt et al., 2000; Oliveira, 1999; Wolfe, 1994).
20
A common approach to studying technological learning in economics is to use a learning curve model, which builds the relationship between learning and experience. Learning curve model describes the phenomenon of learning by doing. Many empirical studies consistently show the robust relationships between firms’ production experience and cost reduction/productivity growth (Arrow, 1962b; Bollinger and Gillingham, 2014; Nemet, 2011; Qiu and Anadon, 2012). Learning curves can be used to estimate future working time and number of laborers. There are two main types of learning curves: the traditional learning curve only includes one factor, cumulative experience; newer models includes two factors and try to separate the effect of R&D and cumulative experience on technological change (Qiu and Anadon, 2012; Söderholm and Klaassen, 2007; Söderholm and Sundqvist, 2007).
Firms can learn from both internal sources or external sources. Figueiredo (2002) classified learning processes into knowledge acquisition and conversion using four categories: 1) external knowledge acquisition; 2) internal knowledge acquisition; 3) knowledge socialization (converting knowledge from individual learning to organizational learning); 4) knowledge codification (converting tacit knowledge to codified knowledge). Researches also identified different types of learning channels. These learning channels can be understood as sub-processes of the above four learning processes. For example, internal knowledge acquisition includes learning by doing, learning by using, learning by searching, and learning by training; external knowledge acquisition includes learning by doing, learning from advances in science and technology, learning from inter- industry spillover, learning by interacting, learning by hiring, learning by alliance, learning by imitating, and learning by training (Arrow, 1962a; Enos, 2008; Hitt et al., 2000; Kline and Rosenberg, 1986; Malerba, 1992; Taylor et al., 2006). Some learning processes (or channels) will be discussed in more detail in Section 2.3 on the theories of geography of innovation, where local networks and learning channels play an important role.
21 Additional factors that can influence a firms’ learning that are explored in the literature include location characteristics, industry characteristics and firm characteristics, such as size, structure, managerial ability, culture, R&D investment, human resources, stock of knowledge, absorption ability, industry structure, external technological information availability, and collaboration (Cohen and Levinthal, 2006; Henderson, 1994; Hitt et al., 2000; Sen and Egelhoff, 2000; Teece, 1993).
WHY AND HOW GEOGRAPHY MATTERS IN INNOVATION
Researchers have recognized that innovation activity and knowledge flows have a tendency to be geographically localized. Although firms are exposed to convenient transportation tools and instant communication tools, geography is still an important element in technological learning and innovation. Over a century ago, Marshall (1895) proposed three reasons why location is important: specialized skilled labor pools, an infrastructure of complementary industries, and proximity to a strong knowledge base. Firms can significantly benefit from “being in such locations”.
Geographic proximity is important for facilitating the spillover of tacit knowledge through face-to-face communication, deep interaction, and even local gossip. The difficulty in codifying tacit knowledge prevents its diffusion to other locations, thus tacit knowledge accumulates in a certain location benefitting the local pool of firms. Tacit knowledge also tends to be contextual, which means that local tacit knowledge is more useful for local firms that are in the same context (institutional context, climate context, etc.) and where personalization to meet user demands are more likely to be similar such that user-producer interaction becomes an important source of technological learning and innovation (Brem and Voigt, 2009). For example, Bathelt and Zeng (2012) studied the chemical industry in Shanghai and found that user-producer interaction significantly contributes to firms’ innovation.
22 Localized learning can also shape local networks, as evidenced by the positive relationship between geographical proximity and closed networks. Geographic proximity facilitates trust- building and enables close collaboration, which is beneficial to localized technological learning. Networks ties developed through interaction with local universities, suppliers, utilities, manufacturers and end-users are all useful to firms’ technological learning and innovation (Grübler, 2003; Helfat, 2006; Junginger et al., 2005; Salter and Laursen, 2006; Schneider et al., 2008). Strupeit and Neij (2017) studied mounting technologies and cost dynamics in Germany, and found that the learning process of mounting technologies has a significant local nature, because local rooftop structure, local building codes and the needs of local installers are significant elements, which means interactions between mounting equipment producers and installers are key in technological learning. Laursen et al., (2012) understood the localized network and norms as social capital and found a positive relationship between geographically bound social capital and firms’ innovative ability.
Geographical proximity also facilitates local firms in forming their practice community. “Community of practice” provides another useful perspective to study localized learning. The literature of “community of practice” is built on collective and interactive learning (Wenger, 1998). Firms working in a similar industry in certain geography can form a community where they share and communicate their knowledge, information, and activities. The connection and interaction enhance firms’ stock of information and knowledge, thus improving firms’ practice (Brown and Duguid, 1991; Wenger, 1998). Another related perspective is from the competition view, where firms can compare their performance to other local firms and reassess, improve or change their practice or performance accordingly (Bathelt et al., 2004).
The view of “community of practice” is consistent with the some key observed features in the solar mounting/racking market. Morrris et al., (2013) compared mounting/racking process between Germany and the U.S. from pre-installation to racking attachment, to module installation,
23 to on-roof electrical and to off-roof electrical and found that practices in the U.S. are different from practices in Germany. For example, in the process of pre-installation, Germany has no off-site preparation that can improve productivity, but in the U.S., off-site preparation is a necessary routine. This indicates that each geographic area has formed their community of practice.
From the perspective of public policy, regional innovation systems are important to facilitate localized learning. Maguire et al. (2012) summarized two policy trends that are inspired by localized learning. The first policy trend focuses on mobilizing local resources and local assets and designing local strategies to promote local innovation and economic growth. The second policy trend is that national innovation strategy began to recognize the importance of regional innovation systems, especially the role of using local resources, local assets and enhancing national innovation policy impacts. Studying regional innovation systems (RIS), especially the nature and characteristics of RIS, includes a focus on firm-level technological learning and networks (Martin and Sunley, 2003; Maskell, 2001).
Although localized knowledge spillover, facilitated by geographical proximity, is efficient for tacit knowledge, localized knowledge has limitations as well. In this vein, a contrasting stream of research contends that non-localized knowledge should also be an important part of technological learning processes. Localized knowledge tends to be more homogenous, thus if we recognize diversity as useful in technological learning and innovation, non-localized knowledge should be expected to be a valuable source for external and diverse inputs (Jacobs, 1969). Asheim and Isaksen (2002) found that too much focus on localized knowledge may lead to path dependence. Geographic proximity makes learning networks fixed and dependent. An effective way to break path dependence is to absorb non-localized knowledge, which tends to be more novel and comprehensive than localized knowledge. Owen-Smith and Powell (2004) studied knowledge network in the Boston Biotechnology Community and found that interaction with “pipeline” (non- localized knowledge) was the source of remarkable innovation. Burt (1992) and Grabher (2002)
24 thought that localized knowledge is highly homogenous and highly stable, thus, extra-local sources have an important role in innovation. Bathelt et al. (2004) built a theoretical framework to illustrate that each firm can benefit from network ties outside local clusters and that information or knowledge obtained from the “global pipeline” can also benefit local clusters through “local buzz”.
Localized learning and non-localized learning have different searching processes. Localized learning is effective by just “being there,” because it is relatively easy to build trust and interact with others in the same geography. As Bathelt et al. (2004) said “localized learning can largely take care of itself”. On the other hand, trust, collaboration, and networks are relatively difficult to establish over long distances, and searching extra-local information and extra-local knowledge requires extra effort, cost, and more complicated processes (Bathelt et al., 2004). However, as noted before, policy trends largely focus on regional innovation systems, such as how to mobilize local resources and help local firms establish collaboration with each other. Therefore, an important policy implication of localized learning versus non-localized learning is to emphasize the necessity of establishing “global pipelines” for local firms and providing more institutional support for non-localized learning.
Some descriptive studies show that choices of learning sources depend on the nature of innovation, the maturity level of industries, and knowledge types. From Schumpeter’s combination view of innovation, researchers believe that diverse knowledge, especially knowledge from external industries and geographically faraway places, is necessary for disruptive innovation (Ahuja and Lampert, 2001; Arthur, 2007; Fleming, 2001; Nemet, 2012a). But incremental innovation may need more specialized knowledge, which may be readily provided by localized learning. Knowledge that is suitable for non-localized learning is codified, because codified knowledge is easy to commute over long distances. Tacit knowledge is difficult to diffuse over long distances, but tacit knowledge can be obtained just by “being there” (geographic proximity), which means that the predominant way of transferring tacit knowledge is through localized
25 learning (Bathelt et al., 2004; Lorenzen, 2007; Storper and Venables, 2004). The maturity level of industries also plays a role in different learning sources. Tacit knowledge is important for industries that are at an early stage. When an industry becomes mature, standardized products and mass production make knowledge easier to diffuse globally (Grossman and Helpman, 1991).
In sum, research on the geography of innovation aims to emphasize the importance of local areas and how local conditions relate to technological innovation. However, from the opposite perspective, densely connected localized clusters may also be treated as barriers for large-scale knowledge spillover.
SOFT COSTS REDUCTIONS AND TECHNOLOGICAL INNOVATIONS IN SOLAR ENERGY
The center of cost reduction literature and technological innovation literature in the field of solar PV has been hardware, panels and modules, in particular. Only very recently have researchers begun to study the dynamics of soft (i.e., non-hardware) costs and cost-reduction opportunities they provide. One motivation for this change in focus is the increasingly important role of soft costs to overall PV system costs. Generally, soft costs include marketing and customer acquisition, system design, installation and engineering costs, permitting and inspection costs, and installer margins. Although solar PV system price has decreased by about 69% between 2000 to 2017 (Barbose et al., 2018), solar PV installed prices are still not competitive with other electricity supply options. Compared to Germany, China, Italy, Japan and Australia, residential and non- residential PV system prices are higher in the U.S., and the disparity in soft costs is the primary reason (Barbose et al., 2015). As it is shown in Figure 4, in the U.S., soft costs account for about 60%-70% of solar PV system prices (Barbose et al., 2017). As a major driver of PV system prices, studying solar soft costs has attracted increased attention to historical trends in soft costs and the factors that influence soft cost reductions, but this is still a small body of literature.
26
Figure 4: Residential Solar PV System Pricing. Source: SEIA/Wood Mackenzie Power & Renewables U.S. Solar Market Insight
Nemet et al., (2016) studied factors that influence the prices of solar PV systems by comparing low-price (LP) systems and non-low-price (non-LP) systems, finding that competition, installer scale, markets, demographics, ownership, policy, and system components are all statistically significant factors. Gillingham et al. (2016) explored factors that influence the variation of solar PV system prices across states in the U.S. and found that federal policies, differences in market structure (i.e., competition), new construction or retrofit, installer experience (learning by doing), inclusion of battery storage and tracking equipment all contribute to price variation in solar PV systems. They also found that the unexplained variation remains large after controlling for all of these variables, which indicates more study is needed to understand price discrepancy that could lead to cost reduction. Both these papers point to the role of experience in reducing soft costs. However, while more experience can help improve labor productivity, it can also lead to technological innovation. Moreover, the process of accumulating experience is also a process of learning from other collaborators. These papers do not separate the effect of labor productivity improvement (i.e., streamlining the installation process), the effect of technological innovation, and the effect of learning from other collaborators.
27 The disparity in soft costs being the primary reason for significantly higher solar PV system prices2 in the U.S. has driven some recent research to study the source of soft cost variation between the U.S. and Germany. Seel et al. (2013) summarized the following factors that may contribute to significantly higher soft costs in the U.S.: smaller market size, higher inverter cost, higher customer acquisition costs, lower customer success rates, higher marketing and advertising costs, longer installation process, higher wages, higher labor hour requirements for permitting, interconnection and inspection (PII), higher fees for PII, and higher sales taxes. Seel et al. (2013) conducted two surveys in the U.S. and Germany and found that the total difference of non- hardware costs for residential PV between Germany and the U.S. is about $2.70/W. Specifically, customer acquisition costs, installation labor costs (including working time and working wage), PII processes, and sales/value-added tax can explain $1.38/W difference. The remaining $1.32/W cannot be explained by the above factors. Strupeit and Neij (2017) pointed out future research should study other factors, especially the role of knowledge, learning process and institutional development.
Comparison studies between the U.S. and Germany support the above research direction. Market growth is more correlated with soft cost reductions in Germany. Country-level cumulative installation can explain 91% of the variation of soft costs in Germany, but only 48% in the U.S (Seel et al., 2013). The learning rate of soft costs in Germany is about 15%, but only 7% in the U.S. (Seel et al., 2013; Strupeit and Neij, 2017). Both are lower than the learning rate for PV modules, which is about 21% on average for global scale studies (de La Tour et al., 2013). Strupeit and Neij (2017) implied the importance of studying and supporting deployment processes after they studied the cost dynamics of solar PV deployment in Germany. The lower learning rate of soft costs and the lower effect of market growth on soft costs also indicate that studies of the
2 The median installed price of customer-owned PV systems that is smaller than10 kW in the U.S. for California Solar Initiative (CSI) was about $5.52/W in 2012Q3, but the system price in Germany is only $2.51/W in 2012Q3 (Seel’s presentation on page 10).
28 learning process relevant to soft costs are important. Such studies could inform policies for supporting learning environments for soft cost reductions.
Studies of soft costs reductions are still a small body of literature. Current literature addresses the role of market size (or market growth) on soft costs reductions to some extent but does not address the separate roles of labor productivity and technological innovation in soft costs reductions. Morrris et al. (2013) explored each step in the mounting/racking process, including pre-installation, racking attachment, module installation, on-roof electrical, and off-roof electrical. They found that German installers can finish each of these steps two to four times quicker and better. Institutional and organizational arrangements, different building structures, and installation practice all contribute to differences in installation efficiency between the U.S. and Germany. The comparison between compositions of soft cost and mounting practices between the U.S. and Germany relates to the geography of innovation addressed Section 2.3, because knowledge about mounting practices and soft costs in Germany (non-localized knowledge) can provide an important learning channel for the U.S. The importance of non-localized knowledge suggests that focusing mainly on localized learning channels or the lack of non-local learning channels may be a barrier to industry development in the U.S. to some extent, and research is needed to resolve the role of non-localized knowledge in the technological learning process.
PATENT DATA AND CITATION ANALYSIS
Patent data is a valuable and available resource for analyzing the process of technological change. In the field of technological change, patent data has been widely used, with various limitations, to evaluate study innovation, its drivers and impacts, including to predict the future development direction of technologies, and to trace knowledge spillover among patents.
Patent counts are widely used to measure the innovative output and technological capacity of a firm, a region, an industry, or a country (Almeida et al., 2002; Griliches, 1990; Hausman et
29 al., 1984; Hottenrott and Lopes-Bento, 2014; Jaffe and Trajtenberg, 2002; Lanjouw and Mody, 1996; Patel and Pavitt, 1987; Scherer, 1983; Schmookler, 1962). However, critics argue that patent counts can only reflect the quantity of technological innovation and may not fully reflect the value or quality of technological innovation. Therefore, deeper indicators, beyond just patent counts, need to be developed from patent data or additional data are needed to reflect the quality of innovation.
Tong and Frame (1994) proposed to use numbers of patent claims as an indicator of technological capacity. They pointed out each claim in a patent can be treated as a separated invention, thus, more claims in a patent indicate a higher value. Patent renewal data (i.e., year of renewal) and patent “family” size can also be used to reflect the value of patents. Patent holders are more likely to pay renewal fees when the economic benefits of patents can cover patent costs (Schankerman and Pakes, 1986). The willingness to file a patent in more countries also indicates that the patent is more valuable (Harhoff et al., 2003; Putnam, 1997). Lanjouw (1998) and Schankerman and Pakes (1985) use patent renewal to build a stochastic renewal model or model of decision to renew that can estimate the rate of innovation. The number of forward citations a patent receives is another widely used indicator for the quality of a patent (Hall, 1999; Trajtenberg, 1990). Forward citations are the number of citations received by a patent. Reitzig (2004) suggested another indicator from the procedure of filing a patent: filing an accelerated examination request during the filing process can indicate the value of a patent.
Griliches et al. (1986) emphasized that forward citations and family size reflect the value of a patent more than any other indicator by comparing all these indicators. Reitzig (2004) summarized and analyzed nine indicators that reflect the economic value of a patent, including lifetime of a patent, patent-breadth, technical non-obviousness, difficulty of “inventing around” a patent, “science linkage” for backward citations, disclosure, portfolio position, uses or functions
30 of patents, and bargaining chips. Which indicators provide the most value depends on the specific industry.
Full-text analysis of patents is increasingly popular, because using the whole abstract or the whole text of a patent is more valuable than simple indicators. Using topic modeling, a computer science technique, can help analyze the full-text of patents by classifying patents, building new measurements for the value of patents, and matching patent treatment groups and patent control groups (Azoulay et al., 2007; Boon and Park, 2005; Kaplan and Keyvan, 2013; Mogoutov and Kahane, 2007; Saiki et al., 2006; UPHAM et al., 2009; Venugopalan and Rai, 2015; Winston Smith and Shah, 2013). Kaplan and Vakili (2013) pointed out that citation analysis can reflect the usefulness of a patent, so they built a new measurement to indicate the novelty of a patent by using topic modeling. Based on Kuhn’s ideas about novelty, patents that introduce new language can be treated as novel, but patents with more forward citations can only be treated as more useful.
Citation data is not only used for determining the value of a patent, but also for tracing knowledge spillover through a patent citation network. Jaffe et al. (1993) is the first paper that uses patent citation to illustrate the importance of localized learning processes. Although Jaffe et al. (1993) faced critiques and questions, using patent citations as a proxy of knowledge flows has been recognized as a valid way to study knowledge spillover. Backward citations of a patent can reflect where knowledge of this patent comes from, while forward citation can be used to trace where the new knowledge goes. Inspired by Jaffe et al. (1993), Caballero and Jaffee (1994), Jaffe and Trajtenberg (1999, 1996) Johnson and Popp (2001), Kelly and Hageman (1999), Sonn and Storper (2003) all use citation analysis to study knowledge spillover. Jaffe and Trajtenberg (1996) and Jaffe and Trajtenberg (1998) focus mainly on citations to university patents and government patents. Kelly and Hageman (1999) studied knowledge clusters by controlling production clusters and found that geographic concentration of knowledge still exists. John et al. (2006) added a time
31 variable into the model and found that knowledge can be spread to larger geographical areas over time.
Current literature about measurements and methodologies show the validation of using patent data and citation analysis. The literature acknowledges the limitations of citation analysis as well. Botolfmaurseth and Verspagen (2002) said, “it should be emphasized that knowledge spillovers are a much broader concept than what is captured by patent citations (US or European)”. For example, citation networks cannot capture tacit knowledge spillover. In addition to patent data and patent citations, researchers also use survey data to study knowledge flows. It is worthy to note that Feldman et al. (2009) verified that geographic proximity is important in knowledge spillover by using survey data. Therefore, if feasible, using both patent data and qualitative methods (i.e., interviews and survey data) is a more robust approach to develop a more holistic, triangulated understanding of knowledge flows.
32 Chapter 3: Local demand-pull policy and energy innovation: Evidence from the solar photovoltaic market in China3
INTRODUCTION
Technological innovation and competitiveness of regional economies are closely connected, driving governments to adopt various policy instruments to support innovation, hoping such support would translate into positive economic effects. A central challenge for governments, though, is to design innovation policies that can effectively influence the speed, direction, and scale of technological innovation. A commonly-used policy tool, especially by sub-national governments, is demand creation. Inspired by Griliches (1957) and Schmookler (1962), the literature started to emphasize that market demand is an important driver for inducing technological innovation. It is now recognized that going from innovation to market demand is not a sequential, linear process, rather it is an interactive, dynamic process whereby the influence (between innovation and demand) can flow both ways.
From a firm’s perspective, the process of innovation involves interactions not only within a firm, but also with other elements outside the firm, such as other firms, consumers, and even the government ((Rothwell, 1989; von Hippel, 1986). Thus, firms’ innovative outcomes cannot be understood well if the environment in which they operate is not taken into account (Antonelli, 1996). This ‘interactive’ view of innovation further raises the question whether local market demand is essential to developing a competitive local industry. Existing literature offers different answers, often at odds with each other, to this question (Dechezleprêtre and Glachant, 2014; Lanjouw and Mody, 1996; Peters et al., 2012; Popp, 2006; Popp et al., 2011).
One area where a strong case for the positive, interactive feedback between geography, innovation, and localized economic effects may be expected involves technologies that have a
3 This chapter is published in Energy Policy. Below is the full citation: Gao, X., & Rai, V. (2019). Local demand- pull policy and energy innovation: Evidence from the solar photovoltaic market in China. Energy Policy, 128, 364- 376. Xue Gao contributed to collecting data, building models, analyzing results, and writing of the manuscript.
33 significant knowledge dependence on local factors such as climate, regulations, building codes, and user preferences. Furthermore, to the extent innovation depends on local knowledge, it may be expected to advantage local firms more, since they presumably have better knowledge of the local context. This possibility should pique the interest of policymakers, who are often interested in seeing that public support for expanding local markets contributes to local technological innovation and economic competitiveness, rather than benefiting other jurisdictions more heavily (Lewis and Wiser, 2007; Maguire et al., 2012). However, empirical evidence in this area is relatively thin. Accordingly, I aim to inform this discussion through robust, empirically-driven analysis. Specifically, using patent data for the distributed solar photovoltaic (PV) market in China from 2005 to 2014, in this chapter I address the following questions: (1) Is there evidence of demand-induced innovation? (2) Does the effect of local demand-pull policy differ from the effect of non-local demand-pull policy on demand-induced innovation? These questions are especially interesting since the essential role of local demand in inducing innovation has been challenged by the fact that China has taken the lead in the PV module manufacturing industry in the absence of a large domestic solar PV market demand when PV manufacturing in China flourished (Huang, 2016; Huenteler et al., 2016; Schmidt and Huenteler, 2016).
This chapter considers solar PV balance-of-system (PV BOS) technologies as a case to address the above research questions. The choice to focus on PV BOS technologies is motivated by two factors: (1) large-scale deployment of distributed solar PV technologies is widely considered to be an important piece in addressing the environmental impacts of the electricity sector, making PV an important technology to study; furthermore, there is a general consensus that the biggest barrier for a wider deployment of PV is the high price of non-module components, including PV BOS (Barbose et al., 2015; Goodrich et al., 2012; Seel et al., 2013; Smith and M. J. Shiao, 2012; Yu et al., 2015), and (2) technological and operational aspects of non-module PV- system components make local conditions (climate, regulation, learning networks, etc.) salient, thus affording a rich empirical setting to explore the geography-innovation-economy interactions
34 at the local level. Although innovations in PV BOS may be generally considered as incremental innovations, they are directly related to solar PV adopters’ experience by making PV installations more user-friendly, safe, and reliable, while also reducing installation costs.
Using original and unique data on innovation in solar PV balance-of-system (PV BOS) technologies in the Chinese distributed solar PV market, this chapter estimates the effect of local demand and non-local demand on technology innovations. This chapter offers two main contributions to the existing literature. First, leveraging a clear demand-pull policy context and careful empirical design, I provide new empirical evidence to reinforce the explanatory ability of the demand-induced hypothesis. A big limitation in previous literature is that many empirical studies do not separate the shift in the demand curve from changes in other conditions (Chidamber, S. und Kon, 1994; Mowery and Rosenberg, 1979; Gregory F. Nemet, 2009), thus confounding the impact of demand on innovation. In this empirical case, the market for distributed PV in China has largely been created by policies, meaning that these market formation policies have led to a significant shift in the demand curve. The clear existence of a policy-driven demand shift can help effectively addresses a key limitation in prior literature. Moreover, strong policy signals, as in this empirical case, mean that installers and manufactures are more readily able to identify and respond to market demand, thus creating the basis for how demand is linked to innovation.
Second, I distinguish between the roles that local demand-pull policy and non-local demand-pull policy play in innovation, thereby highlighting the importance of local demand in driving locally-generated innovation. In particular, I exploit variation in province-level PV BOS patenting activities in China to address the influence of locational aspects of demand on innovation. Most empirical studies that have explored this locational aspect have focused at the national level (i.e., domestic vs. foreign), thus my use of subnational policy and demand variation adds a new perspective to the literature on this topic. I also exploit differences in technological characteristics within PV BOS to explore how the linkage between location of demand and innovation depends
35 upon the characteristics of the involved technologies. Thus my analysis also offers insights into the potential for tailoring market-creation policies based on technological characteristics.
BACKGROUND AND RELATED LITERATURE
Related literature
Scholars from a range of disciplines have increasingly emphasized the role of technological innovation as an engine of economic growth (Schumpeter, 1934). Thus understanding the drivers of and barriers to technological innovation has excited much interest across a range of disciplines for several decades.
The induced innovation hypothesis, proposed by Schmookler and Griliches decades ago, emphasizes that the power of induced innovation lies in market demand (Griliches, 1957; Schmookler, 1962). Utterback further pointed out that market factors have a primary influence on innovation and that 60% to 80% of “important innovations” are demand-induced in a large number of fields (Utterback, 1974). Many recent empirical studies also provide strong evidence to support the demand-induced innovation hypothesis, namely that innovation can be induced by demand
(Lanjouw and Mody, 1996; Lin, 2010; Gregory F. Nemet, 2009; Peters et al., 2012; Popp, 2006b). It is noteworthy that demand-induced innovation is not a simple linear model. Feedback loops and interactions between different actors – between firms, between users and firms, and between governments and firms – are also important elements of technological innovation in the context of demand-induced innovation (Kline and Rosenberg, 1986; Rothwell, 1989; Taylor, 2008; von Hippel, 1986).
A parallel stream of literature has explored the linkages between interactive learning and innovation (Christensen and Stoerring, 2011; Lundvall, 2012; Rothwell, 1989; von Hippel, 1986). Learning and innovation often result from an interactive process, which benefits from proximity, thus experience-based innovation is often characterized by localization and geographical
36 agglomeration. The importance of local context factors (such as demand, interaction, culture, regulation, business environment, and supply chain) in the innovation process has been recognized (Asheim et al., 2007, 2011; Fabrizio and Thomas, 2012; Jaffe and Trajtenberg, 2002; Marshall, 1895; Nemet, 2012; Porter and Stern, 2001). Firms with more experience in a certain geographic area have comparative advantages because of their better understanding of the local context (Beise and Rammer, 2006; Fabrizio and Thomas, 2012; Kogut, 1995) and are stronger for business-to- consumer sectors, as they focus more on end-users and demand-induced innovation in those geographic areas (Brem and Voigt, 2009). Thus, it is plausible that demand-pull policy could expand local markets and help firms gain more localized experience, translating into (demand- induced) innovation. However, policymakers often prefer to see that public support for expanding local markets contributes to local technological innovation and economic competitiveness, rather than benefiting outside jurisdictions (Lewis and Wiser, 2007; OECD, 2011).
Recently some empirical work has begun to distinguish between the effects of local demand-pull policy and non-local demand-pull policy on technology innovation, suggesting substantial variation in these effects depending on the industry/sector of study (See Table 1). For example, Popp (2006) found that innovations for sulfur dioxide and nitrogen oxide control at coal- fired power plants primarily respond to domestic regulation instead of foreign regulation. Dechezleprêtre and Glachant (2014) found that technological improvements in wind power across OECD countries respond to both domestic and foreign demand-pull policies, although they also suggested that the marginal effect of domestic policies is twelve times greater than that of foreign policies. In contrast, Peters et al. (2012) found that both domestic demand-pull policies and foreign demand-pull policies could spur innovation in the solar PV (modules) sector, with similar magnitudes of effect for both foreign and domestic demand-pull policies. On the other hand, Lanjouw and Mody (1996) found that vehicle emissions regulation in the U.S. stimulated innovation in Japan and Germany. Popp et al. (2011) observed that technology innovation in the pulp and paper industry responded to both domestic and foreign regulations.
37
Table 1: Summary of relevant empirical studies to test whether the location of demand matters for inducing innovation.
Industry/Technology Domestic Policy Foreign Literature Policy Sulfur dioxide and nitrogen √ Popp (2006) oxide control Wind technological √ √ Dechezleprêtre and improvements Twelve times greater Glachant (2014) √ Solar PV manufacturing √ Peters et al. (2012) Similar magnitude
Vehicle emissions √ Lanjouw and Mody (1996) Pulp and paper industry √ √ Popp et al. (2011) Note:A tick mark denotes if a study found domestic and/or foreign demand-pull policy to significantly induce (domestic) innovation. The text in the ‘Domestic Policy’ column denotes the relative impact on innovation of domestic vs. foreign demand pull.
The variation in these results regarding the impact of policy locus on demand-induced innovation may depend on whether an industry faces a global market (Popp et al., 2011). For example, the U.S. is a major market for Japanese vehicle makers; thus Japanese technology development is appropriately responsive to emissions regulation in the U.S. The seemingly inconsistent results may also depend on whether local context is significantly important to technology innovations (Peters et al., 2012). If technology localization and producer-user interaction are important in an innovation process, such experience-based innovation might depend on a stable and sizable local market (Qiu and Anadon, 2012). This chapter focuses on PV BOS technologies, for which local factors are known to be more significant than they are to PV modules. Innovations in PV BOS are a product of, among other things, local experience and local learning, as well as local climate, local rooftop characteristics, and local complementary industries (Neij et al., 2017; Venugopalan and Rai, 2015). Therefore, based on existing literature, I expect that PV BOS innovations should primarily respond to local demand-pull policies.
38 Background on distributed PV policies in China
Since 2005 China’s solar PV manufacturing industry has developed rapidly. In 2007 Chinese solar PV manufacturing surpassed Japan, and China became the world's largest producer of PV modules. However, in 2011 the Chinese PV industry slowed down significantly due to shrinking foreign markets, Europe in particular. To help the Chinese PV industry digest overcapacity and get through the downturn, Chinese governments at various level expended much effort to expand the domestic market.
Since 2011 the Chinese national government has established several demand-pull policies to expand the distributed PV market, including a grid-connection policy, Feed-in-Tariff policy, generation subsidy policy, and financial services policy (National Development Reform Commission, 2013; National Grid Company, 2012; State Council, 2013; Zhang, 2016). The deployment of distributed PV in China is also through the following programs: the Golden Sun Demonstration Projects jointly led by China’s Minister of Finance, China's Ministry of Science & Technology and National Energy Agent (Ministry of Finance, 2009); the Green Building Plan led by China’s Ministry of Housing and Urban-Rural Construction (Ministry of Housing and Urban- Rural Construction, 2013; National Development Reform Commission, 2013); the distributed- generation PV application and demonstration projects led by National Energy Agent, and individual self-consumption distributed-generation PV.
39
5000 4500 4000 3500 3000 2500 (MW) 2000 1500 1000
lnstalled capacity of distributed PV PV distributed of capacity lnstalled 500 0 2011 2012 2013 2014
Figure 5: Installed capacity (MW) of distributed PV4 in China, 2011-2014. Data source: NEA (2014), NEA (2015) and Zhang (2016).
At the subnational levels, provincial, city, and local governments made great efforts to expand local market sizes. By the end of 2014, 11 provinces issued feed-in tariff policy and other supplementary subsidies to support solar PV deployment. For example, Jiangxi Province offered the feed-in tariff level at 0.2 yuan/kWh and provincial-specific funding is at 4 yuan/W and 3 yuan/W with different eligibility conditions. Some feed-in tariff policies were combined with local content requirement policies. For example, Shangluo City in Shaanxi province provided 0.1 yuan/W and 0.05 yuan/kWh for solar PV projects using local PV products. It also provided favorable tax policies, specifically a 5% local tax refund to locally registered balance-of-system firms. City-level governments also issued demand-pull policies. Hefei City at the Anhui Province provided additional city-level 0.25 yuan/kWh subsidy for roofs and building integrated solar PV for 15 years and 2 yuan/W for residential PV projects. For most provinces, only one or two cities
4 Distributed PV systems include projects with electricity generation located near users with total installed capacity no greater than 6 MW; the electricity is mainly consumed by end users, but excess electricity can be fed into the public grid (National Development Reform Commission, 2013; National Energy Agency, 2014; National Grid Company, 2012).
40 provided additional subsidies at the city level. But in Zhejiang province, twelve local governments offered feed-in tariffs to local markets, and these city-level subsidies can be combined with national and provincial subsidies.
In addition to feed-in tariffs and other subsidies, provincial and local governments also made great efforts in helping find proper rooftops, clarify rooftop ownership, finance, and install and interconnect distributed PV systems, several provinces, especially the Eastern provinces, provide detailed grid-connection services and financing services to local distributed-generation PV projects. For example, the provincial government in Jiangsu assesses all available and suitable rooftop resources for distributed PV and helps address the problem of unclear rooftop ownership. The provincial government in Jiangsu also requests Grid Company to provide specific services to help costumers connect distributed PV electricity to the grid, so that consumers can sell their excess electricity back to Grid Company. Owing to such coordinated government support, distributed PV has developed rapidly in the Eastern provinces of China. Literature shows that these demand-side policies in China are the major divers of solar PV deployment (Corwin and Johnson, 2019; He and Liu, 2014; Hou et al., 2019; Urban et al., 2018; Zhang et al., 2015; Zhang, 2016; Zhao et al., 2015). Therefore, these demand-side policies and developments yield helpful variation (for my analysis) in the level of demand-pull policies and distributed PV deployment across the Chinese provinces.
Increasing installation of distributed PV in China has significantly expanded the practical experience of PV installers with installing and maintaining distributed PV systems. As discussed above, incremental innovations, such as those in PV BOS, are expected to be closely related to local interactions, local customer preferences, local regulation, local business environment, and local supply chain (Bollinger and Gillingham, 2014; Fabrizio and Thomas, 2012; Maskell, 1999; Neij et al., 2017; Nemet, 2012b; Porter and Stern, 2001). In turn, local firms have comparative advantages that result from a better understanding of the local context (Beise and Rammer, 2006). With expanding markets, installers have more opportunity to learn about the nature of local
41 demand, local roof structures, and local codes and regulations. Such practical experiences provide increased opportunities for installers to improve and update their PV BOS related products and services.
DATA AND METHODOLOGY
Data
To explore the effect of local demand on PV BOS innovation, I build an original database of PV BOS patents filed between 2005 and 2014 related to distributed PV based on patent information from the State Intellectual Property Office (SIPO) website. SIPO contains invention announcements, invention authorization, utility models, and design patents. After SIPO receives an application for an invention, it conducts a preliminary review. If the application passes the preliminary examination, the invention is typically announced 18 months after the filing date. After an invention application is announced, SIPO starts a substantive examination if requested by the applicant. If the application passes the substantive examination, it is authorized to be an invention. There is no substantive examination for utility models and design patents. A utility model is similar to an invention, but standards of novelty and inventiveness for utility models are not as high as those of inventions. A design model refers to new designs for a product’s appearance, shape, and color.
China patent system was established in 1985, marked by the passage of the first patent law. Then Chinese patent law experienced four amendments in 1992, in 2000, in 2008, and in 2016, separately. The biggest change in the Chinese patent system happened in 2000. As China prepared to join the World Trade Organization (WTO), they were required to fulfill the Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS). While my descriptive analysis covers the Chinese PV BOS patents during 2005 and 2014, there are only few patents before 2008. Thus, the change of China patent system in 2008 should not have a significant impact on my descriptive analysis. Furthermore, my regression analysis covers the PV BOS patents from 2012
42 to 2014, during which there was no significant change in the Chinese patent system. Thus, patent data should be a stable indicator during my study period in regression analysis.
For identifying PV BOS patents for each province in China, I adapt the methodology developed by Venugopalan and Rai (2015), which focused on PV BOS patents in the U.S. The first step is to identify specific items and keywords for PV BOS. Following Venugopalan and Rai (2015), this chapter classifies BOS into four categories: inverter, mounting system, monitoring, and site assessment. Inverter is used to convert power generated by solar PV cells/modules from direct current (DC) to alternating current (AC). Mounting structures are to physically support solar PV modules and/or arrays. Monitoring includes methods and devices that monitor health, connectivity and output of solar PV systems. Site assessment includes methods or mechanisms of appraising the photovoltaic output potential of a land parcel, determining the ideal placement of PV systems on potential sites, determining the extent to which obstructions or shade may impede the output potential of PV sites (Venugopalan and Rai, 2015). All these non-module components are used in the solar PV deployment process. In other words, solar panel/module is used to convert solar energy to DC; however, in order to complete a whole deployment process, non-module components (i.e., PV BOS) are needed for identifying the placement of PV modules, attaching PV modules to rooftops, converting DC to AC, and monitoring the operation and performance of PV modules.
Specific items I include for each category can be found in the appendix A. I use the same search terms as Venugopalan and Rai (2015) in English and translate these English terms into Chinese. Note that one keyword in English could be translated into several relevant terms in Chinese. Table 2 lists keywords I used for searching relevant patents of PV BOS.
43 Table 2: Search terms for PV BOS patents in Chinese and English.
BOS category Search Terms Translation in English
Inverter “逆变器” and “光伏”or “太阳能” "inverter" and "photovoltaic" or "solar"
“支架” or “框架”or “齿条”or“屋顶”or "mounting" or "bracket" or "frame" or Mounting “安装”or “围栏”or “轨道” and “光伏”or "roof" or "installation" or "rail" and “太阳能” "photovoltaic" or "solar"
“跟踪” or “评价” or “诊断” or “维持” or "monitoring" or "evaluation" or Monitoring “保持” or “监视” or “监视器” or “维护” "diagnosis" or "maintain" and or “监控” and “光伏”or “太阳能” "photovoltaic" or "solar" “卫星” or “地图” or “阴影” or “遮蔽” or "satellite" or "map" or "shade" or “估计” or “预测” or “支持” or "estimation" or "projection" or "support" Site “安装”and “光伏”or “太阳能” or "installation" and "solar" or Assessment "photovoltaic"
The second step is to use keywords for each category to search patents’ full text on the SIPO website. This step gives me a large sample of potentially relevant patents, but not all patents I extract from keyword searching are actually relevant to PV BOS. After the second step, my search process yields a total of 7626 potentially relevant patents. The third step is the key step, wherein I review each patent’s claims and identify whether the invention embodied in a patent is directly related to one of the four PV BOS categories. Note that a patent may have been an application in earlier years, but went on to become a granted patent in the following years. I remove these overlapping patents. I also drop the patents that are related only to utility-scale PV, since the distributed and utility-scale markets are largely distinct and innovations in utility-scale PV are not expected to be induced by the demand for distributed PV. In any case, the “utility-scale only” patents that I remove account for a very small proportion of total potentially relevant patents.
After the review in the third step, I retain 2122 patents filed between 2005 to 2014 as being actually relevant to PV BOS innovation, and this set forms the basis of the analysis reported below. The fourth step is to extract the date and the location (province) of each patent. A patent has three
44 important dates: filed date, application announcement date, and invention authorized date. Filing date is the better choice for measuring when an innovation actually took place (Griliches, 1990). I use location of the patentee to measure where an innovation took place. In this study, I include all 31 provinces, municipalities, and autonomous regions in China.
Measurement
I use the number of patents in PV BOS as the dependent variable to measure technology innovation in PV BOS. Though with limitations, patenting activity is frequently used as a meaningful proxy for measuring innovative activity. Numerous studies recognize that patent count is an appropriate and useful measure of innovative activity (Almeida et al., 2002; Choi et al., 2011; Griliches, 1990; Hausman et al., 1984; Jaffe and Trajtenberg, 2002; Lanjouw and Mody, 1996; Schmookler, 1962; Venugopalan and Rai, 2015). Patent data is also widely used in the literature as a measure of innovation in China (Choi et al., 2011; Choi and Williams, 2014; Dang and Motohashi, 2015; Sun and Du, 2010; Wang et al., 2015).
I acknowledge limitations of using patents as a measure of innovation. The number of patents does not map directly on to innovation (Botolfmaurseth and Verspagen, 2002). For example, implicit knowledge cannot be captured by patents. Likewise, not all firms are willing to protect their innovations by applying for patents, because they may not be willing to publicly disclose their ideas (Cohen et al., 2000). While there is no perfect measurement for innovation, given that patent data offer the most systematically compiled, spatially and temporally nuanced data about inventive activity in China (Choi et al., 2011) and the steps I take to check robustness of the results, I believe that patenting activity is an appropriate measure for the purposes of this study.
It is problematic to define the scope of demand-pull policy and categorize specific innovation-support policy as demand-pull policy or not (Kemp and Pontoglio, 2011; Taylor, 2008).
45 Furthermore, the major demand-pull policies (i.e., feed-in tariff policy and subsidies to initial investments) vary a lot across provinces and cities with respect to level, duration, eligibility, and other features, and it is very difficult to aggregate different ways of city-level demand-side subsidies with different level, duration, and eligibility at the provincial level (see details on the background section). Local governments’ efforts in finding proper rooftops, clarifying rooftop ownership, financing, and installing and interconnecting distributed PV systems are even harder to quantify. A commonly-used measure of demand-pull policy in the PV industry is PV installed capacity (Dechezleprêtre and Glachant, 2014; Klaassen et al., 2005; Peters et al., 2012; Rai et al., 2013). In the case of distributed PV market in China, literature shows that market demand in China has mainly been created by government policies, especially at the early stage of solar PV deployment (Corwin and Johnson, 2019; He and Liu, 2014; Hou et al., 2019; Urban et al., 2018; Zhang et al., 2015; Zhang, 2016; Zhao et al., 2015). Accordingly, I believe it is appropriate to use province-level distributed PV installed capacity as a proxy for demand-pull policy. The limitation of this measurement choice is that only three years of data are available in China, because the Chinese government began to promote the distributed PV industry and collect relevant data only since 2011. The limited availability of installed capacity data constrains my modeling choices. I discuss those limitations and how I address them in more detail below. Specific description of the measurement for each independent variable is provided in Section 3.3.
Since innovative activity is influenced by provincial-level innovation capacity and innovation environment (Cohen and Levinthal, 1990; Cooke et al., 1997; Lundvall, 2012; OECD, 2011; Porter and Stern, 2001; Prajogo and Ahmed, 2006), these factors need to be accounted for in my analysis. Provincial-level innovation capacity in China varies a lot (Sun and Liu, 2010; Yang and Lin, 2012). Innovative provinces might be more likely to engage in PV BOS innovation activities; thus, I use total number of patents in each province as a control for overall innovative capacity. The incentives for patenting likely vary across provinces, as provinces may implement various pro-patent policies and/or have variations in accessing the patenting system (Li, 2011). A
46 province with more favorable patent promotion policies is likely to have more patenting activity. Thus the general propensity to patent may also be controlled by the total number of all patents in a province. In addition, the development of complementary sectors could also influence PV BOS innovation activities (Choi and Anadón, 2014; Pisano and Teece, 2007). I use earnings of four sectors including manufacturing, utility, wholesale & retail, and scientific & technical services as controls for development of complementary sectors. Table A4 in appendix A summarizes descriptive statistics of key variables and Table A5 shows the correlation table between the independent variables.
Empirical models
Since the number of patents is a type of count data, two appropriate count models are Poisson regression and negative binominal regression. Additionally, the number of patents for each province in this chapter has over-dispersion. Since the mean and variance of a Poisson regression are the same, Poisson regression cannot be used for data with over-dispersion. However, a negative binominal distribution has two parameters, one of which is an over-dispersion parameter, thus negative binominal regression is a better choice for count data with over-dispersion. To check the robustness of the regression results, I also include the regression results of using Poisson regressions in appendix A.
In my database, variation of provincial level installed distributed PV within province over time only accounts for about 9% of the total variation in provincial level demand, and variation of provincial level installed distributed PV between provinces accounts for 91% of the total variation. Variation of annual added installed capacity within province over three years (2012-2014) may not be enough for an efficient two-way fixed effect model. When variation across provinces is great but variation within each province over time is little, fixed effect estimates would be imprecise and standard errors are likely to be too large to tolerate (Allison, 2009). In this case, a random effect model that takes advantage of both within-province variation and between-province
47 variation may be a better choice. Therefore, I use the Hausman Test to compare a random effect model and a fixed effect model. I estimated the following model by using both fixed effect estimation and random effect estimation:
������ = (u + � ) + β ln ����� + β ln �������� + β ln � + � , (1) ������ = u + β ln ����� + β ln �������� + β ln � + (� + � ), (2) ������ = (u + � ) + β ln ����� + β ln �������� + β ln � + � , (3)
where the dependent variable ������ is the number of PV BOS patents for province i in year t; ln ����� is logged local installed capacity (i.e., for a given province) of distributed PV for province i in year t; ln �������� is logged non-local installed capacity (i.e., all of China minus a given province) of distributed PV for province i in year t, and ln � is a vector of log of time variant control variables (fully listed in table A4 of appendix A) for province i in year t. u is time fixed effects, which could be different for each year. α is the unobserved time-invariant individual effect, which only varies among provinces. � is an error term. This error term is different for each province in each year. For the first fixed effect model (Equation 1), α and u are considered as a part of the intercept, so it allows correlation between α and the predictors. In a random effect model (Equation 2), α is considered part of the error term, so it does not allow any correlation between α and the predictors.
I expect that differences in learning processes exist between subcategories of PV BOS technologies (inverter, monitoring, mounting, and site assessment). Inverter technologies are expected to have the least local nature, compared to the other three subcategories, because inverters are closer to globally-traded products that are relatively easily adaptable to the local context. Therefore, I split the PV BOS patents into inverter and non-inverter (to include monitoring, mounting, and site assessment) patents, meaning that I use the total number of inverter and non- inverter patents as two separate dependent variables to run the regression models. The differentiation between inverter and non-inverter patents allows me to explore the technological
48 dependence of whether or not local demand is a prerequisite for inducing localized innovation. Huenteler et al. (2016), Schmidt and Huenteler (2016), and Quitzow et al. (2017) found that whether or not firms in developing countries without a large home market are able to catch up with developed economies in certain sectors depends on the nature of associated learning processes, suggesting the criticality of the location of demand in the development of comparative advantage. This provides the basis for my inverter/non-inverter split models. I expect that the effect of non- local demand on inverter patents will be much larger than the effect on non-inverter patents.
The inverter/non-inverter splitting serves another important purpose. While I use the total number of patents in a province to control for potential differences in propensity to patent across provinces, the potential differences in propensity to patent across BOS categories should also be taken into account. In principle, there are two reasons (Basberg, 1987; Fontana et al., 2009) that potentially cause differences in patenting propensity across BOS categories. First, it may be hard to codify an innovation as a patent. In the case of PV BOS, inverter technologies are most likely to be codified. But non-inverter technologies, such as installation methods, mounting skills, and personalized design, contain more tacit knowledge. As such they are relatively harder, and thus also less likely, to be codified. As a result, an inverter-related innovation is more likely to be filed for a patent. Second, inventors may believe that their innovations are not innovative enough to be granted patent, leading them to forgo attempting to patent, given the non-trivial effort and costs associated with the patenting process. In the case of PV BOS, non-inverter technologies are more likely to be incremental innovations, compared to inverter technologies, and thus are less likely to be filed as patents. Considering these two reasons, in order to account for the potential differences in propensity to patent across BOS categories, it is appropriate to split patents into inverter and non-inverter patents.
To check whether the regression results are sensitive to the quality of patent data, I conduct robustness checks by including only ‘inventions’, which are known to be of the highest quality
49 among different types of patents in China. In other words, I redefine the dependent variable as the annual number of PV BOS inventions for each province from 2012 to 2014, thus excluding utility models and design patents when constructing the dependent variable. Another purpose of this robustness check is to compare the strength of the effects of local and non-local demand on all patents versus only inventions (high-quality patents).
Finally, in order to further address the limitation posed by the relatively small sample size (n = 90) and the small variation in annual added installed capacity within the provinces over three years (2012-2014), I also develop an alternative model. Since the two-way fixed effect model including province fixed effect and year fixed effect will discard information (variation) across provinces and thus lead to large standard errors, in the alternative model I group provinces and estimate a two-way fixed effect model including year fixed effect and group fixed effect (Equation
3). In Equation 3, α is the unobserved time-invariant group effect. I use three ways to group provinces. The first approach is based on geography, where I group provinces into seven groups: Eastern provinces, Northern provinces, Western provinces, Southern provinces, Central provinces, Northeastern provinces, and Southwestern provinces. The provinces in the same geographic areas share similar climate, culture, business environment, and economic growth conditions. The second approach is based on Gross Domestic Product (GDP). I rank the provinces according to provincial GDP in 2014 and group provinces into six groups according to the ranking: top 5 richest provinces, top 6-10 provinces, top 11-15 provinces, top 16-20 provinces, top 21-25 provinces, and top 26-31 provinces. The third approach is based on a comprehensive competitiveness index according to the province ranking in the “Blue book of China’s provincial competitiveness” (Li et al., 2015). I also group provinces into six groups according to the ranking. Figure A1 to Figure A3 in the appendix A illustrate the three ways to group 31 provinces in China.
50 Omitted variables and potential endogeneity
Omitted variable bias potentially exists, because location-specific variables and institution- relevant variables, such as social value, culture, norms, customs, and institutional structures, are potential confounding variables. However, these confounding variables typically change rather slowly, so I do not expect omitting them to impact the results from the models, which focus only on a three-year period. As such, since these confounding variables may be treated as time-invariant during my study period, they would be adequately controlled for through the province-level fixed- effects in my models.
Reverse causality – PVBOS innovations driving demand over the three-year study period – is not a major concern in this study, because the Chinese government’s support for expanding the domestic PV market since 2011 is driven by a shrinking foreign market, Europe in particular. The effort in creating domestic market demand is to help the Chinese PV industry digest overcapacity and get through the downturn. Put differently, market demand (the independent variable) in this case is genuinely exogenous variation, which is not a result of the development of innovations in the field of PV BOS (the dependent variable) in China.
RESULTS
Descriptive analysis of PV BOS patents in China
The Chinese PV BOS patent dataset illustrates trends for PV BOS innovation activities in China from 2005 to 2014. Figure 6 shows the number of PV BOS patents in China over time. The number of PV BOS patents increased from 2 in 2005 to 567 in 2014. A big jump in the number of PV BOS patents occurred in 2011, increasing by over 100% from 155 in 2010 to 326 in 2011, which is the year the Chinese government began to change policy direction from large-scale PV to distributed PV deployment. Under the Chinese government’s strong support, installations of distributed PV in China increased from 0.6GW in 2011 to 2.3GW in 2012. Utility patents account for a higher proportion of the total number of patents than inventions, which indicates that PV
51 BOS patenting activity in China has tilted more toward incremental improvements rather than disruptive innovation.
Design 600 Utility 500 Invention
400 Application
300
200
100
Annual nmuber of patents (counts) 0 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014
Figure 6: Annual distribution of PV BOS applications, inventions, utility and design (counts).
Figure 7 shows distribution of annual number of patents by category. The overall trend for all four categories is increasing from 2005 to 2014. The relatively flat number of inverter patents since 2012 mainly results from the decreasing number of design patents filed (which refer to new designs for a product’s appearance, shape, and color). The number of monitoring patents and site assessment patents have increased slowly but steadily since 2011. Patenting activity related to mounting is more relevant to the expansion of local markets, since better understanding of local context, such as buildings and rooftops, provides comparative advantage for local firms.
52 350 5000 Inverter 4500 300 Monitoring 4000 Mounting 250 3500 Site Assessment 3000 200 2500 150 2000 100 1500 1000 50 500 Annual number of patents 0 0 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 Distributed PV installed capacity (MW)
Figure 7: Annual number of PV BOS patents including invention application, invention, utility and design (count), by category and installed capacity of installed distributed PV (MW). Note: The left axis is the number of PV BOS patents (count) and the right axis is installed capacity of installed distributed PV (MW).
The geographic distribution of PV BOS patenting activity correlates with the geographic distribution of distributed PV installations in China. Figure 8 and figure 9 show the geographic distribution of both distributed PV installations and number of PV BOS patents. Nine provinces in China have few PV BOS patents during 2012-2014, which coincides with almost zero distributed PV installation in these provinces; these provinces mainly focus on utility-scale PV systems. Five provinces have more than 100 PV BOS patents in China during 2012-2014, three of which (Jiangsu, Guangdong, and Zhejiang) are the top three provinces with highest cumulative distributed PV installations. These richer provinces suffer from frequent electricity shortages, but land-intensive utility-scale PV conflicts with their scarce and expensive land. As a result, distributed PV is a reasonable and necessary choice for these rich Eastern provinces.
53
Figure 8: Geographic distribution of the provinces with less than or equal to 10 PV BOS patents from 2012 to 2014 (blue) and the provinces with more than or equal to 60 PV BOS patents from 2012 to 2014 (red).
Figure 9: Geographic distribution of the provinces with nearly zero MW installed capacity of distributed PV from 2012 to 2014 (blue) and the provinces with more than or equal to 500 MW installed capacity of distributed PV (red).
54 Regression model results
Regression results for Equations 1 and 2 are presented in Table 3. Column 1 shows the regression results of the two-way fixed effect model. I am interested in the effect of local demand and non-local demand on PV BOS innovative activity. The results show that local demand has a positive relationship with local PV BOS patenting activity, but that non-local demand has a negative relationship with local PV BOS innovation. Neither of these two coefficients is statistically significant, which is consistent with my previous expectation that only 9% of variation within provinces over three years in my database could lead to large standard errors. Column 2 shows the regression results of the random effect model. The random effect model takes advantage of variation between provinces and within provinces over time. The result of the Hausman test shows that the coefficients of the two-way fixed effect model (model 1) and the random effect model (model 2) are consistent, but estimates of the random effect model (model 2) are more efficient, which means the random effect model is a more appropriate model. The results of the random effect model show that the effect of local demand for distributed PV in stimulating local PV BOS innovation is positive and statistically significant.
The regression results shown in column 3 of Table 3 try to follow the approach used in Peters et al. (2012). The methodology used in Peters et al. (2012) includes an aggregated time trend variable to reflect three development phases of PV in Europe. For Europe, Peters et al. (2012) defines 1974-1985 as the first time-trend, 1986 to 1994 as the second time-trend, and 1995 to 2009 as the third time-trend, and thus use time-trend fixed effect instead of year fixed effect, which would be much more granular (as is the case with model 1 and model 2 in Table 3). In my case, since (for modeling purposes) I have data for only three years (2012-2014) and all these three years may be taken to be part of the first boom of distributed PV development in China, I take this to mean that I have only one time-trend. Thus, in effect, this means that I do not need a dummy variable for this time trend. Therefore, model 3 includes only the province fixed effect. The results of model 3 are similar to those of model 2. Both models show that local demand (i.e., within a
55 province) can positively stimulate local PV BOS innovations. As expected, the total number of patents in each province is also statistically significant. The total number of patents reflects the overall innovative capacity of each province, which is positively related to the number of PV BOS patents.
Table 3: Regression results for Equations 1 and 2 using different estimation methods.
Model 1 Model 2 Model 3 Fixed effect Random effect Fixed effect 0.121 0.208*** 0.244** Log local demand (MW) (0.127) (0.0770) (0.0987) -3.096 -2.426 -0.893 Log non-local demand (MW) (2.069) (1.879) (0.578) 0.536* 0.505*** 0.453* Log total patent (counts) (0.275) (0.113) (0.257) 0.868 -0.420 0.466 ¥ Log manufacture salary ( ) (2.223) (1.420) (2.107) 0.770 -0.116 0.748 Log utility salary (¥) (0.785) (0.568) (0.800) 2.435* 1.615 2.041 ¥ Log wholesale & Retail salary ( ) (1.336) (1.077) (1.344) Log scientific & technical services -0.657 0.332 -0.957 salary (¥) (0.933) (0.630) (0.944) -14.33 0.454 -19.67* Constant (23.54) (15.64) (11.95) Year fixed effect Yes Yes No Province Fixed effect Yes No Yes Observations 90 90 90 *** p<0.01, ** p<0.05, * p<0.1
Table 4 shows the regression results of the province grouping approach to address the problem that variation within province over time only accounts for 9% of total variation in province-level annual installed capacity. As discussed in data and methodology section, I apply three ways to group provinces and include both year fixed effect and group fixed effect in these three models (Equation 3). Models 4, 5, and 6 (Table 4) show regression results for the model grouping provinces based on province ranking in the “Blue Book of China’s Provincial Competitiveness (2015),” geography, and GDP, respectively. The coefficients of local demand in all three models are positive and statistically significant, ranging from 0.17 to 0.26. The magnitude
56 and standard error of the coefficients of local demand in model 4 and model 6 are consistent with the results in model 2 and model 3. All these models show that only the local demand for distributed PV significantly promotes local PV BOS innovation, and that non-local demand for distributed PV does not affect local PV BOS innovation.
Table 4: Regression results for different grouping methods (Equation 3), based on competitiveness index according to the city ranking in the “Blue book of China’s provincial competitiveness,” geography, and GDP.
Model 4 Model 5 Model 6 Fixed Effects Fixed Effects Fixed Effects Group by index Group by geography Group by GDP 0.257** 0.165** 0.241** Log local demand (MW) (0.107) (0.0757) (0.0952) 1.010 -2.703 -2.847 Log non-local demand (MW) (3.189) (2.379) (3.367) 0.626*** 0.748*** 0.666*** Log total patent (Counts) (0.156) (0.0963) (0.164) 0.271 -1.829* -0.432 Log manufacture salary (¥) (1.400) (1.028) (1.515) -0.161 -0.103 0.00903 Log utility salary (¥) (0.634) (0.610) (0.732) 0.224 0.391 0.00388 Log wholesale & Retail salary (¥) (1.309) (0.976) (1.511) 0.784 1.039* 1.207 Log scientific, & technical services salary (¥) (1.007) (0.553) (0.766) -24.47 18.87 7.743 Constant (24.07) (20.20) (29.10) Year fixed effect Yes Yes Yes
Group Fixed effect Yes Yes Yes
Observations 90 90 90 *** p<0.01, ** p<0.05, * p<0.1.
Table 5 shows the regression results for models where I split the patents into inverter and non-inverter patents. The dependent variables in models 7, 8, and 9 are, respectively, the annual number of all BOS patents at the province level, the annual number of inverter patents at the
57 province level, and the annual number of non-inverter patents at the province level. The coefficients of local demand are consistent with the results of model 1 to model 6 (Tables 3 and 4), further confirming that local demand has an essential role in spurring local innovative activity in PV BOS. While the coefficients of non-local demand are not significant, likely due to limited variation in the non-local demand variable, the opposite signs of the coefficient estimates on inverter and non-inverter patents offer interesting insights. The coefficient for non-local demand on inverter patents is positive, while it is negative on non-inverter patents. This indicates, as expected, that innovations in inverters – a technology that is more globally traded than any of the other BOS categories considered here – could potentially learn from non-local markets. This also helps explain why non-local demand seems to have little or no impact on BOS innovations but positively impact innovations in PV modules. Essentially, previous literature suggests that innovation in PV modules is a phenomenon of global learning, but innovation in PV BOS, on the other hand, is strongly driven by local learning (Neij et al., 2017; Shum and Watanabe, 2008; Wene, 2000; Wiser et al., 2007). Since local learning is predominantly driven by local demand, local learning appears to be the linkage between the positive impact of local demand on PV BOS technologies. In other words, these results suggest that whether or not a local market is a precondition for inducing local innovative activity depends on technology characteristics: local market has an essential role for technologies that have a significant local nature and, thus, depend on local learning for innovations in them.
58 Table 5: Regression results for differentiation between inverter and non-inverter patents. Note that Model 7 is the same as Model 2.
Model 7 Model 8 Model 9
Dependent Variable All BOS patents Inverter patents Non-inverter inventions
Log local demand (MW) 0.208*** 0.267** 0.217** (0.0770) (0.105) (0.0910) Log non-local demand (MW) -2.426 0.998 -3.091 (1.879) (2.826) (2.104) Log total patent (counts) 0.505*** 0.727*** 0.445*** (0.113) (0.147) (0.127) Log manufacture salary (¥) -0.420 1.854 -2.724 (1.420) (1.904) (1.684) Log utility salary (¥) -0.116 -1.444** 0.945 (0.568) (0.732) (0.672) Log wholesale & Retail salary (¥) 1.615 -0.755 2.904** (1.077) (1.414) (1.360) Log scientific, & technical services 0.332 0.260 0.871 salary (¥) (0.630) (0.875) (0.739) 0.452 -14.21 -0.0751 Constant (15.64) (22.61) (17.61) Year fixed effect yes yes yes Observations 90 90 90 *** p<0.01, ** p<0.05, * p<0.1.
Table 6 shows the regression results for models where the dependent variable includes only patent inventions, which are known to be of the highest quality among different types of patents in China. To be specific, the dependent variable in models 10, 11, and 12 are the annual number of all inventions at the province-level, the annual number of inverter inventions at the province- level, and the annual number of non-inverter inventions at the province-level, respectively. These regression results are consistent with previous models (Tables 4 and 5). As before, only local demand has a significant and positive effect on local innovations. Results from these models also indicate that the strength of the effect of local demand on BOS inventions (high-quality patents) (model 10) is somewhat larger than the effect on all BOS patents (model 7). Most interestingly, as
59 expected, the coefficient of local demand on non-inverter inventions (model 12) is much larger compared to inverter inventions (model 11), which may indicate that local demand has a stronger positive effect on non-inverter inventions compared to inverter inventions. This finding needs to be tempered, though, since the coefficient of local demand on non-inverter inventions (model 12) is significant only at the 10% level. Overall, these results suggest that local demand might play a critical role in inducing high-quality local patents in PV BOS. This intuitively makes sense, since larger demand offers greater opportunity to benefit from innovations – and the magnitude of those benefits, in turn, will be larger for higher-quality patents, thus providing a multiplicative incentive.
Table 6: Regression results for including only inventions in the dependent variable.
Model 10 Model 11 Model 12
All BOS Non-inverter Dependent Variable Inverter inventions inventions inventions 0.249** 0.274** 0.391* Log local demand (MW) (0.120) (0.136) (0.221) 1.183 5.704 -3.591 Log non-local demand (MW) (3.289) (3.906) (4.253) 0.700*** 0.907*** 0.564** Log total patent (counts) (0.162) (0.191) (0.236) -0.0484 2.729 -2.804 Log manufacture salary (¥) (2.114) (2.173) (3.556) -0.607 -2.161*** 1.157 Log utility salary (¥) (0.835) (0.749) (1.360) 0.812 -0.657 2.229 Log wholesale & Retail salary (¥) (1.729) (1.728) (2.604) Log scientific, & technical services 0.131 -0.518 1.053 salary (¥) (1.002) (0.863) (1.299) -19.41 -48.18 18.78 Constant (26.28) (30.52) (1012.7) Year fixed effect yes yes yes Observations 90 90 90 *** p<0.01, ** p<0.05, * p<0.1.
60 DISCUSSION AND LIMITATIONS
The results show that for PV BOS in China only local demand significantly induces local PV BOS innovation, while non-local demand has no effect. This emphasizes that local demand could translate into local innovative activity. But why should local PV BOS innovation be spurred only by local market experience (i.e., demand)? I suggest that learning by doing and learning from geographical agglomeration are the two main mechanisms through which such a translation might happen.
On one hand, with expanding local markets, installers have more opportunity to learn about local factors relevant to the design and installation of a PV system, such as local roof structures, codes and standards, supply-chain, and the nature of consumer demand – which, for novel technologies such as PV is expected to have market-specific heterogeneity, especially in early phases of the market. Such local installation experiences provide increased opportunities for installers to improve and update their PV BOS related products and services. On the other hand, with increasing local demand, knowledge gained by each firm operating in the local market contributes to the overall local pool of knowledge. A stronger local knowledge base, in turn, also improves local firms’ innovative potential. Such knowledge is often described as if it were in the air, and local firms benefit from local knowledge just because of “being there” (i.e., without having to expend much effort). Moreover, from the view of competition, in order to compete better in the local market, local competition pushes firms in the local cluster to improve their practices and innovations (Bathelt et al., 2004).
However, the perspective of geographic agglomeration also highlights an inherent limitation of this chapter: my analysis does not fully capture the benefits associated with tacit knowledge that are gained from local clusters, because inventive activities that are captured by patent data mainly embody codified knowledge. Geographic agglomeration is expected to be more important for tacit knowledge, as the spillover of tacit knowledge significantly benefits from face-
61 to-face communication, deep interaction, and even local gossip, and the difficulty in codifying tacit knowledge prevents its diffusion to other locations (Gertler, 2003; Haldin‐Herrgard, 2000; Venkitachalam and Busch, 2012). Thus, using only patenting activity as a measure of innovative outcome in this chapter means that this chapter likely underestimate the effect of local demand on local innovations in PV BOS.
It is worth noting three other limitations of my analysis and how I have tried to address them. First, as noted before, patenting activity is not necessarily a good measure of technological innovation. Previous literature particularly points to concerns about the quality of Chinese patents compared to that of U.S. or European patents. In this chapter, because the timeframe of my analysis is very recent (2012-2014) to effectively detect forward citations, I am unable to weight patent count by using forward citations as a way of addressing the quality of patents. Instead, I conduct a robustness check by excluding utility models and design patents – known to be lower quality patents in China – from my models in order to check whether my regression results are sensitive to the quality of patent data. Dropping lower quality patents does not change the main findings. A second noteworthy limitation of my analysis is that fixed-effects models have limited ability to infer causal relationships, compared to conducting random experiments. Previous literature shows that confounding variables, especially institutional factors and various patenting incentives across provinces, should be carefully controlled for in studies using China patent data. In this chapter, I do not include data on provincial-level institutional factors. However, I believe that these institutional factors, such as social value, culture, norms, customs, and institutional structures, are slow-changing variables. As such, these factors may be expected to not change significantly within three years (2012-2014, the focus of my analysis). Thus, I include province fixed-effects in my models to control for these (time invariant) variables. Thirdly, PV installed capacity is a commonly-used proxy measure for demand-pull policy in the PV industry (Dechezleprêtre and Glachant, 2014; Peters et al., 2012; Rai et al., 2013). This is partly because it is problematic to define the scope of demand-pull policy and categorize specific innovation-support policies as
62 demand-pull policy or not (Kemp and Pontoglio, 2011; Taylor, 2008). Moreover, the major demand-pull policies, namely feed-in tariff policy and subsidy to the initial investment, vary a lot across provinces and cities with different level, duration, and eligibility. It is even harder to quantify local governments’ efforts in finding proper rooftops, clarifying rooftop ownership, financing, and installing and interconnecting distributed PV systems. Yet, I acknowledge that using distributed PV installed capacity at the province level as a proxy for local demand pull is not the most direct way to measure demand-pull policies. Future work could try to address this limitation by leveraging the variation of specific demand-pull policies province-by-province to improve measurement of the local demand-pull variable.
CONCLUSION AND POLICY IMPLICATIONS
In this chapter I have addressed the following question: is local demand essential in inducing local innovation? my empirical design exploits the subnational variation in (policy- created) demand to study the impact of geographically-differential (local vs. non-local) demand on local innovation. Specifically, I use a new patent dataset for photovoltaic balance-of-system technologies at the province level in China to study the effect of local demand and non-local demand on province-level PV BOS innovation.
Geographic distribution of distributed PV installations is unequal across the 31 provinces in China, but it is strongly correlated with geographic distribution of PV BOS patents. The consistency between the explosion of PV BOS patents and local demand is especially telling in the case of mounting technologies: to enable adequate physical support over the life of the PV system (20-30 years), mounting equipment need to be designed for local climatic conditions and variations therein, rooftop structures, and framing and roofing materials. Thus, greater local demand is expected to strongly drive innovations in mounting technologies to cater to that local demand, as I indeed find to be the case.
63 I find that only local demand for distributed PV significantly contributes to local PV BOS innovation in China. This suggests that local distributed PV markets created by demand-pull policies (of both provincial and local governments) significantly promotes (only) local innovative activity in PV BOS. The effect of non-local demand on patenting activity in inverters, while insignificant, is large and positive, but is negative on patenting activity in non-inverter BOS technologies. This suggests, consistent with expectations, that innovations in inverters may benefit from non-local markets, given the more fungible nature of inverter technologies. As expected, overall innovative capacity at the provincial level – as measured by all (i.e., not just PV BOS) patenting activity within each province – significantly and positively affects PV BOS innovation. The significantly different effects of local demand and non-local demand indicate that local context (including local knowledge and local experience) is a key element in the learning and innovation process for PV BOS. The importance of local knowledge and local context in demand-induced hypothesis has a valuable market implication: to the extent innovation depends on local knowledge, it advantages local firms more, since they presumably have better knowledge of the local context. In such cases local demand-pull policies may be more politically feasible, since policymakers would view the potential economic benefits from market creation to be more appropriable by local actors.
Furthermore, my findings serve to inform that drivers of innovation depend on market and technology characteristics. PV BOS helps emphasize that within the same technology (an integrated PV system, here) the technological trajectories of different subcomponents (modules, inverters, etc.) potentially interact differently with local market conditions. Previous studies have suggested that different learning processes are likely at play for PV cell/module versus PV BOS (Neij et al., 2017; Shum and Watanabe, 2008; Wene, 2000; Wiser et al., 2007). PV module innovation is a phenomenon of global learning, and there is no clear boundary between local PV module learning and global PV module learning. However, PV BOS innovation is strongly driven by local learning, which comes from accumulated learning of installations in specific, local
64 markets (Neij et al., 2017). The differences in technology characteristics and the associated learning processes help explain the apparent inconsistencies of my findings with the results of Peters et al. (2012). Peters et al. (2012) analyzed innovation that is directly related to PV cells and modules. They noted that user feedback and producer-user interaction are weak for PV cells and modules – a situation that does not hold for PV BOS technologies, the focus of my study.
Moreover, differences in learning processes are not limited to just between PV cell/module and PV BOS technologies; rather, those differences exist even between subcategories of PV BOS technologies (inverter, monitoring, mounting, and site assessment). Indeed, I find that for the highest-quality patents (i.e., inventions in the Chinese patent system) local demand has a stronger positive effect in inducing innovation in non-inverter BOS technologies compared to inverters. This is consistent with the fact that mounting structures and site assessment technologies have the most prominent local nature. In the case of inverters, on the other hand, although local demand is still a significant factor in spurring innovation, I also find weak support that inverter innovations in China may have been driven by non-local markets. Overall, these findings suggest that whether local demand is an essential precondition to local industry development (i.e., innovations) depends on technological characteristics and the nature of associated learning processes, highlighting the need to open the “technology black box”. In other words, a better understanding of the impact of market-creation policies on the geography innovation necessitates accounting for technology characteristics and in particular how those characteristics accentuate certain learning processes over others.
My findings also insightfully inform the discussion on ‘green industry development’. The ultimate goal of green industrial policies is to integrate multiple policy goals including, but not limited to, environmental protection, energy security, technological innovation, and economic competitiveness (Rodrik, 2014; Schmidt and Huenteler, 2016). My results suggest that demand- pull policies for the distributed PV market in China have significantly shaped local innovative
65 activities. However, the important role of local demand-pull policy in inducing innovation has been challenged by the fact that China has taken the lead in the PV module manufacturing industry even in the absence of a large domestic solar PV market demand when PV manufacturing in China flourished (Huang, 2016; Huenteler et al., 2016; Schmidt and Huenteler, 2016). These arguments have raised a critical question to the literature on green industrial policy: is local demand essential in the development of green industry? My study suggests that the answer to this question depends, at least in part, on technology-specific factors. Specifically, by comparing technology subcategories within PV BOS as well as comparing PV BOS technologies with PV module, I find that technology characteristics are critical factors in explaining the role of local demand in green industry development. If learning and innovation processes depend on local context, then local demand is a prerequisite for developing a local competitive industry. Moreover, most empirical studies that have explored this locational aspect have focused at the national level (i.e., domestic vs. foreign), thus the use of sub-national policy and demand variation adds a new perspective to the literature on this topic. Importantly, my study hints that the arguments typically made in the discussion regarding green industry development may also be applicable to sub-national levels of policy-making. Demand-pull policy could be an important strategic element in the local policy mix of developing green industry, particularly, for technologies that have a significant local nature. By the same token, it also might be difficult for lagging provinces/states to close the technology innovation gap without developing a stable local market demand. This perspective of looking at the sub-national level is especially important for large countries, such as China, but it may be less relevant to small countries.
The success of subnational governments’ efforts in expanding market demand and translating it into local innovative activities underscores the potential viability of a bottom-up approach to renewable energy governance, at least for some technologies. The localized economic gains from expanding local market demand could mobilize a large number of subnational actors in support of advancing certain renewable energy technologies locally. Thus, if the central/federal
66 government is slow to move on environmental policies, or only focuses on supporting export-led growth, subnational policymakers could still take voluntary, interest-driven actions to expand the market demand for renewable energy, because they would view the potential economic benefits from market creation as more appropriable by local actors. These subnational actions could be voluntary and interest-driven, thus potentially underlying a viable mechanism for the bottom-up approach, even if not coordinated.
Finally, one fruitful direction for further work could be development of a framework to explain the conditions under which local demand is a prerequisite for developing a competitive local industry. These conditions include, but are not limited to, technology characteristics, the degree of market maturity, the development stage of technology, local policies, and local institutional ecosystem. This broader framework can help generalize the results of solar PV BOS to other technologies. A broader framework that situates the role of local demand in inducing local innovation could further help explain relevant mechanisms of localized learning.
67 Chapter 4: Does Geography Matter? A Study of Technological Learning and Innovation in the Solar Photovoltaic Balance-of-Systems (PV BOS) Industry
INTRODUCTION
The importance of geography in technological learning and innovation is an important component of endogenous growth theory, innovation theory, and industrial economics. However, the importance of geography as an element of technological learning continues to be debated, within the underlying context that modern transportation and communications infrastructure – key enablers of deepening globalization – make long distance no longer a barrier in technological learning and innovation. One situation where there is presumably a strong case for the positive, interactive feedback between geography and innovation involves technologies whose future development trajectory significantly depends on local factors (e.g., climate, regulations, administrative codes, user preferences, etc.). Technologies in the deployment process are an example of such a situation, as the features of such technologies are associated with the locations in which they are applied and deployed. In such cases Schaeffer et al. (2004) pointed out that “effective learning only takes place when both deployment is increased and the process of how deployment leads to learning is better understood”. Previous literature only offers a limited understanding of the linkages between the geography of learning and innovation outcomes for technologies in the deployment process – a key research gap in the literature (Schaeffer et al., 2004; Strupeit and Neij, 2017). The salience of this topic is ever greater today, when governments across the world are struggling to balance the opposing pulls of inward-looking, protective approaches to economic growth and outward-looking, collaborative, and open approaches to economic growth. Speaking to these theoretical, empirical and policy opportunities, this chapter focuses on learning and innovation processes of green technologies in the deployment process.
The subject of my empirical inquiry is distributed solar photovoltaic (PV) systems. Large- scale deployment of distributed solar PV technologies is widely considered to be an important
68 piece in addressing the environmental impacts of the electricity sector, making PV an important technology to study. Over the last few years solar PV has been the fastest growing electricity generation technology globally, accounting for about 47% of all newly added renewable energy installations in 2016 (REN 21, 2017). However, PV installed costs need to come down further as solar PV enters the next phase of much broader deployment. The hardware for a solar PV system is typically clubbed in two broad categories: modules – a collection of cells that convert solar irradiation on direct current (DC) and non-module or “balance-of-system” hardware – all hardware in a PV system other than the PV modules. Most of the extant literature about solar energy learning and innovation focus on PV modules, and only several research has begun to pay attention to non- module hardware (i.e., PV balance-of-system) innovation and non-module costs (Neij et al., 2017; Nemet et al., 2016; Shum and Watanabe, 2008; Strupeit and Neij, 2017).
This small body of literature emphasizes the technological learning and innovation processes of non-module hardware greatly differ from innovations of PV module. Shum and Watanabe (2008) identify significant differences, driven primarily by the associated geography of learning processes, when comparing the technological innovation processes of PV module and non-module (balance-of-system) hardware. Technological learning in PV modules is widely considered to be a global phenomenon. However, PV balance-of-system (PV BOS) technologies, which enable the deployment of PV modules in a specific local context (say, a rooftop), typically embody local characteristics (Shum and Watanabe, 2008; Strupeit and Neij, 2017) associated with the local deployment context (e.g., local climate, local rooftop structure, and local regulations and codes, etc.). Thus, localized learning may be expected to be critical for innovations in PV BOS technologies. However, there is very little empirical literature exploring the extent to which localized learning is relevant for PV BOS innovations and the comparative contribution of local vs. non-local knowledge in value added to PV BOS innovations. As a result, studying the technological learning and innovation process of PV BOS contributes to completing the whole picture of technological innovations in the solar PV industry.
69
To address this research gap, in this chapter I empirically study the role of geography in the technological innovation of non-module solar energy technologies (PV BOS) using patent citation analysis on a unique database of PV BOS patents in the U.S. issued between 2000 to 2014. My analysis of PV BOS technologies addresses three questions: 1) how does geographic distance influence knowledge spillover, 2) how technological characteristics influence knowledge acquisition choices as regards local knowledge vs. non-local knowledge? and 3) how do local and non-local knowledge affect the value (or quality) of innovation? The remaining discussion is structured as follows: Section 2 provides theoretical background on geography of innovation and technological learning literature; Section 3 describes data collection and methodology used in this chapter; Section 4 presents results and relevant discussions; and Section 5 summarizes the main insights that emerge from my results, describes the limitations of this chapter, and discusses relevant policy implications.
THEORETICAL BACKGROUND
Studies of geography of innovation have recognized that innovation activity and knowledge flows are geographically localized. Although firms have access to convenient transportation and instant communication tools, geography is still an important element in technological learning and innovation. Because of specialized skilled labor pools, an infrastructure of complementary industries, and proximity to a strong knowledge base that facilitates knowledge exchange, firms can significantly benefit from locating in certain geographic regions (Marshall, 1890).
The role of geography is also an important, but understudied, topic for technological learning and innovation in PV BOS. Shum and Watanabe (2008) raised several key questions about whether technological learning and innovation of PV BOS is local or global in character. Their findings support the notion that PV BOS learning (innovation) spillover is local. Neij et al. (2017)
70 reinforced these finding and identified the need for local learning in solar energy technology deployment. They also described what is learned locally, who learns and how, and how spatial proximity supports learning. However, these descriptive studies cannot quantify that to what extent the localized learning is important to PV BOS learning and innovation.
Geographic proximity facilitates knowledge spillover and network-building, and can also shape local networks (Bottazzi et al., 2002; Gertler, 2003; Jaffe et al., 1993), as evidenced by the positive relationship between geographical proximity and closed networks. Trust and close collaboration are easier to build in the same geographic area. Networks developed through interactions with local universities, suppliers, utilities, manufacturers and end-users are useful to firms’ technological learning and innovation (Grübler, 2003; Helfat, 2006; Junginger et al., 2005; Salter and Laursen, 2006; Schneider et al., 2008). Laursen et al., (2012) understood the localized network and norms as social capital and found a positive relationship between geographically bound social capital and firms’ innovative ability. In the field of PV BOS, Strupeit and Neij (2017) studied mounting technologies and cost dynamics in Germany, finding that mounting technology learning processes are highly local in nature. The authors argue that local rooftop structure, local building codes and the needs of local residents are significant explanatory factors behind the highly local nature of learning processes, and that interactions between mounting equipment producers and installers are key in technological learning. Based on prior literature, I expect that PV BOS innovation, especially mounting and site assessment innovation, will absorb more knowledge from the local geographic area. Alternatively, this hypothesis indicates that geographic distance might be a barrier to knowledge spillover and firms at peripheral regions might have insufficient local knowledge, which may negatively impact local firms’ innovation activities. Relatedly, it is of interest to explore whether accessing non-local knowledge is able to compensate insufficient local knowledge. These illustrations of the limited, but emerging, literature around how geography impacts learning and innovation in PV BOS highlight the need to better understand how local and
71 non-local knowledge influence technological innovation, and whether non-local knowledge can compensate for insufficient local knowledge.
Geographical proximity also facilitates local firms in forming their practice community. Collective and interactive learning are key concepts in the “community of practice” literature (Wenger, 1998). Firms working in the same geographic area and similar industries can form a community where they share knowledge, information, and activities. These industry connections and interaction enhance the stock of information and knowledge firms possess, improving the firms’ practice (Brown and Duguid, 1991; Wenger, 1998). The view of “community of practice” literature is consistent with the operations in the solar mounting/racking market. Morrris et al., (2013) compared the mounting/racking process between Germany and the U.S., including pre- installation, racking attachment, module installation, on-roof electrical and off-roof electrical, and found that practices in the U.S. are different from practices in Germany. For example, during the pre-installation process, U.S. firms routinely conduct off-site preparation that can improve productivity, while Germany firms do not conduct off-site preparation. This process illustrates how geographically rooted communities of practice have formed in relation to PV BOS technologies.
Localized learning and the significance of geography indicate the importance of local knowledge in innovative activities. Local knowledge tends to be more closely aligned with local firms’ technological demand (Arza, 2010) and easier to understand and absorb (Qiu et al., 2017), because they are created and applied in the same context. Generally, local knowledge tends to be more homogenous (Jacobs, 1969). Asheim and Isaksen (2002) found that too much focus on local knowledge may lead to path dependence, and geographic proximity makes learning networks fixed and dependent. An effective way to break path dependence is to absorb non-local knowledge, which tends to be more novel and comprehensive than local knowledge (Asheim and Isaksen, 2002). Thus, knowledge at different geographic levels should have different strengths in
72 supporting innovation. Following Schumpeter, innovation researchers believe that diverse knowledge, especially knowledge from external industries and geographically distant places, is a valuable source for innovation, especially disruptive innovation (Jacobs, 1969; Nelson and Winter, 1982; Ahuja and Lampert, 2001; Arthur, 2007; Fleming, 2001; Nemet, 2012). Owen-Smith and Powell (2004) studied knowledge network in the Boston Biotechnology Community and found that interaction with the “pipeline” (non-local knowledge) was the source of remarkable innovation. Burt (1992) and Grabher (2002) pointed out that local knowledge is highly homogenous and highly stable, thus extra-local sources have an important role in innovation. Bathelt et al. (2004) built a theoretical framework to illustrate that each firm can benefit from network ties outside local clusters and information or knowledge obtained from the “global pipeline” can also benefit local clusters through “local buzz”.
As noted above, technologies associated with solar PV deployment embody various locational characteristics, emphasizing the importance of localized learning in developing valuable and relevant innovations. Thus, I hypothesize that the knowledge acquisition when conducting PV BOS innovations has a significant local dependence, which means that the degree of knowledge acquisition decreases with geographic distance. Within the field of PV BOS, technologies regarding mounting and site assessment are hypothesized to have a higher degree of local dependence when acquiring knowledge and conducting innovation, because compared to technologies regarding inverter and monitoring, those technologies are more related with local context. Although local knowledge is expected to be more relevant in conducting PV BOS innovation, due to different advantages and disadvantages of knowledge in various geographic areas and the importance of knowledge diversity on innovation, I hypothesize that non-local knowledge will be an important driver of quality in PV BOS innovations.
The searching processes for localized learning and non-localized learning differ, which has valuable policy implications. Since it is relatively easy to build trust and interact with others in the
73 same geographic area, localized learning results just by “being there.” Bathelt et al. (2004) go so far as to state “localized learning can largely take care of itself”. On the other hand, trust, collaboration, and networks are relatively difficult to establish over long distances, and searching extra-local information and knowledge requires more effort, cost, and complicated processes than local information and knowledge (Bathelt et al., 2004). Moreover, absorbing non-local knowledge also requires firms to have high absorptive capacity (Qiu et al., 2017). Absorbing non-local knowledge might be a more serious impediment for firms that locate outside the technological core areas, have low absorptive capacity, and are in industries that face greater technological uncertainty (Levy and Murnane, 2004; Sonn and Storper, 2003).
Overall, the discussion above implies that it is necessary and valuable to understand different roles of local and non-local knowledge in technological innovation when policymakers design and prioritize innovation policies. If local knowledge and networks are essential for technological innovation, in order to promote technological innovations, policymakers may want to focus on how to mobilize local resources and help local firms establish collaboration with each other; if non-local knowledge and networks also play important roles in technological innovation, policymakers could further help local firms establish “global pipelines” and provide more institutional support for non-localized learning.
DATA AND METHODOLOGY
Data
I used keyword search strings to develop an original database of PV BOS patents in the distributed PV market from the United States Patent and Trademark Office (USPTO) website, registered between 2000 and 2014. I include four technological areas in PV BOS, including inverter, monitoring, mounting, and site-assessment. Inverters convert power generated by solar PV cells and modules from direct current (DC) to alternating current (AC). Mounting is used to physically support solar PV modules and/or arrays. Monitoring includes methods and devices that
74 monitor health, connectivity and output of solar PV systems. Site assessment includes methods and mechanisms for appraising the photovoltaic output potential of a land parcel, determining the ideal placement of PV systems on potential sites, and the extent to which obstructions or shade may impede the output potential of PV sites. The specific technologies included within each of the four main BOS components can be found in the Supplemental information (SI).
The data collection steps are as followings. The first step is to identify specific items and keywords for PV BOS (See table 7). The second step is to use keywords for each category to search patents’ full text on the USPTO website. This step gives me a large sample of potentially relevant patents, but not all patents I extract from keyword searching are actually relevant to PV BOS. Next, the third step is to review all these patents’ claims and identify whether the invention embodied in a patent is directly related to one of the four PV BOS categories. Further, the same patent may have been an application in earlier years but went on to become a granted patent in the following 2 to 3 years. I remove these overlapping patents for my descriptive analysis. The fourth step is to extract the assignee, the date and the location of each patent. The data collection steps are the same as Gao and Rai (2019).
Table 7: The specific items and keywords for PV BOS patent data collection BOS category Key words Inverter "inverter" and "photovoltaic" or "solar" Mounting "mounting” and "photovoltaic" or "solar" Monitoring "monitoring" or "monitor" and "photovoltaic" or "solar" "satellite" or "map" or "shade" or "estimation" or "obstruction" or "potential" Site Assessment or "site" or "terrain" or "assess" or "azimuth" or "survey" and "solar" or "photovoltaic"
I further used USPTO application programming interface (API) to collect backward citations and forward citations for each PV BOS patent in the database, and for each backward citation and forward citation, my database includes the application year, the location of an assignee, and the longitude/latitude of the location of an assignee. Citations indicate that there are intellectual
75 property claims the new patent yields to. Backward citations are patents that are cited by a patent, while forward citations are patents that cite the focal patent. Taking “patent A” as an example, backward citations of the patent A are patents cited by the patent A, and forward citations of patent A are patents that cite the patent A in the future. By exploring the knowledge flows from backward citations to the focal patent then to the forward citations (the whole knowledge spillover chain), this chapter is able to study firms’ technological learning and information acquisition behaviors. The link between a focal patent and each backward citation can also be understood in the context of the underlying citation network; as such, I assess the effect of geographic distance on the citation networks (i.e., any pairs between PV BOS patents and backward citations).
Methodology
Gravity model
Jaffe et al. (1993) was the first paper to use patent citations to illustrate the importance of localized learning processes. They compare an original database to a control database using t-tests, finding a significant localization effect, that patents are more likely to cite patents that come from the same country and state. Following Jaffe et al. (1993), using patent citations as a proxy for knowledge flows has been recognized as a valid way to study knowledge spillover (Caballero and Jaffee, 1994; Jaffe and Trajtenberg, 1999, 1996; Johnson and Popp, 2001; Kelly and Hageman, 1999; Sonn and Storper, 2003; Johnson et al., 2006; Nemet and Johnson, 2012; Nemet 2012). Backward citations are used to indicate the knowledge sources of the focal patent, and the number of forward citations indicates the value (or quality) of the focal patent (Abrams and Popadak, 2013; Gambardella et al., 2008; Lanjouw and Schankerman, 2004; Schoenmakers and Duysters, 2010).
The first research question in this study asks how geographic distance influences knowledge spillover. I use patent citations to capture knowledge spillover and gravity model to quantify the impact of geographic distance on knowledge spillover. Equation (1) and (2) are the
76 corresponding empirical models. The dependent variable is the number of linkages between the location (the location is a state in the U.S.) of the assignee of a U.S. PV BOS patent and the location of the assignee of all the backward citations in the U.S. PV BOS patent database at a certain year.
In other words, the dependent variable � is the number of citations that the “location a” cites the “location b” at year t, which indicates the scale (number of citation linkages) of knowledge flows from location b to location a. I built all possible pairs between locations of U.S. PV BOS patents and locations of all the backward citations of U.S. PV BOS patents, thus the dependent variable is count data with “excess zero”. Therefore, I apply zero-inflated negative binominal regression model (ZINB) to estimate the model.
Equation (1) is derived from the gravity model, which is one of the most successful empirical models in international economics (Frankel and Rose, 2002). The gravity model is widely used to predict international trade flows that is proportional to economic sizes (or populations) of importer and exporter, and is inversely proportional to the distance between importer and exporter. In the citation-based network, firms in a “location a” can be treated as knowledge importers and firms in a “location b” can be treated as knowledge exporters. Following Morescalchi et al. (2015), the total number of linkages that are attached to a “location a” indicates the size of “location a”, the total number of networks that are attached to a “location b” indicates the size of “location b”. The variable “�������� ” indicates the geographic distance between location a and location b; the variable “state” is a dummy variable, and if state equals 1 means that two regions of this network are in the same state in the U.S.; and the variable “border” is also a dummy variable that indicates whether two regions of this network are both in the U.S. I also include year dummies to control for time trends, and include dummies that indicate the country of “location b” to control for unobserved characteristics of cited countries, especially any unobserved variables that affect the underlying propensity of being a location of any given backward citation. In the equation 2, I include one more control variable, which is used to control for a set of alternative potential cited locations in the knowledge acquisition process using a proxy index
77 called “remoteness”. Head (2003) provides the formula to calculate the index of “remoteness” that measures a location’s average weighted distance from all the potential knowledge spillover partners.
Equation (1)
� = � + � ∗ log(���� ) + � ∗ log(���� ) + � ∗ log(�������� ) + � ∗ ����� + � ∗