The short-term costs of local content requirements in the Indian solar auctions

Model

An ideal experimental set-up to study the effect of LCRs on bid prices would include (at least) two identical countries, completely independent of each other, with an auction scheme identical apart from the LCR feature. Any price difference that emerged between the two auction schemes could then be fully attributed to the differences in the LCR feature. Even better would be to include additional identical countries with varying levels of stringency in the level of the LCRs (in weight, value or number of components), to study whether there are discontinuities in bid prices that seem to be attributable to increasing levels of LCR stringency. For instance, one might expect a non-linear increase in bid prices if very high levels of LCRs (say, 95%) were introduced given that the manufacturing capabilities for wafers would be near zero in India³⁵.

The Indian case comes close to an ideal policy experiment given that many tendered capacities were distributed between LCR and non-LCR auctions in similar geographic areas—albeit not always equally in terms of capacity. Importantly, firms were free to bid in both LCR and open auctions, and many firms submitted bids in both auction windows. However, to address the possible remaining issue of selection bias (that is, firms self-selecting into auctions with or without LCRs), we divide firms into two groups: firms that only bid in auctions without LCRs (59 firms, 134 bids) and firms that bid in both auction types, open and closed (26 firms, 143 bids). The data used in the study come from various sources from the Indian government and firm-level data from Mergent Intellect (described in more detail in the Data section).

We considered using a multinomial selection model that would divide firms into three groups: those that only bid in auctions with LCRs, those that only bid in auctions without LCRs and those that bid in both auction types. However, since there are only seven firms in the category of ‘only LCR auctions’, we were unable to run the model. Yet we believe that the main question is whether a firm bid in an LCR auction, because it indicates whether the firm has sufficient local knowledge to either liaise with a local manufacturer or to use its own existing manufacturing facilities. Firms that only bid in the open auction, in contrast, could merely import the required parts. Hence, we expect there to be systematic differences between these two groups.

Thus, we test whether firms that do not bid in the LCR category are different from those that do bid in the LCR category. It could be, for instance, that firms that bid in LCR auctions have more experience in local development than firms that only bid in the open auctions (where there are no restrictions in using imported material). Similarly, firms that only bid in the open auctions might be able to more effectively exploit economies of scale by producing solar PV cells and modules for several markets (for example, Canadian Solar, which has manufacturing capabilities in China).

In addition to using standard ordinary least squares regressions, we therefore make use of a Heckman regression model, which accounts for this possible selection bias. Heckman’s 1979 seminal paper proposes a two-step statistical approach³⁷. In the first step, an economic model is defined in which plausible factors for the probability of falling into (in our case) either Group 1 or 2 are considered. This is modelled as a probit regression,

where G indicates whether the firm belongs to Group 1 (G = 0 otherwise), Z is a vector of explanatory variables, b is a vector of unknown parameters and Φ is the cumulative distribution function of the standard normal distribution. The explanatory variables we consider are the number of employees of a given firm, whether it is a state-owned enterprise (SOE) and whether the firm is itself a manufacturer or is merely a project developer (an indication of the degree of vertical integration). We also consider whether the firm is primarily focused on energy or merely attempts to diversify from an unrelated field, indicating limited technical experience, and whether the company already bid in the NSM Phase I. The latter factor captures advantages that firms might have in the NSM Phase II due to prior experience with the auction system.

The second stage of the Heckman model then uses the probability that a firm will self-select into Group 1, based on its characteristics, by including that probability as an explanatory variable in the ordinary least squares regression.

For our standard ordinary least squares and Heckman regression model, we also created a number of explanatory variables that we assume influence bid price. We recognize that competition differed substantially between rounds and was on average twice as high in open auctions as in LCR auctions (as measured by our variable defined in equation (2)). Firms are likely to anticipate, or at least have beliefs about, the level of competition in an upcoming auction round, which leads them to adapt their bids accordingly (that is, to make higher bids when less competition is expected; this is well documented in the literature³⁸). In order to exclude the possibility that bids under LCR regulation are higher solely due to this effect, we control for the degree of competition within each round. Therefore, we define the competition for each tender as follows:

where B is the capacity in MW of each of the bids submitted for a particular auction round r, AC is the total capacity in MW auctioned in round r and N is the overall number of bids received for each auction round r. For instance, if 20 MW are auctioned off and firms submit 100 MW in bids, the competition would be 5.

We also include the cumulative installed capacity of each developer within the auction windows we cover to account for learning-by-doing of the developers and capacity building (for example, through greater local knowledge and connection to suppliers)^39,40. Our time dummy controls for exogenous technological change, such as decrease in the cost of solar PV modules and other equipment over time, that is not directly related to the deployment in India (that is, exogenous technical progress)⁴¹. We do not include a state dummy as the variable is correlated too strongly with our time dummy (as certain states only conducted auctions in specific years, leading to high multicollinearity). We do, however, include the mean solar irradiation (annual average kWh m⁻² d⁻¹) per state to control for differences in the solar resources across different states (we also use the maximum solar irradiation for each state as a robustness check, which does not affect the results⁴²).

In addition, we include a dummy for the utility that purchases the electricity generated by the awarded projects. It is well documented that the financial solvency of the utility buying the electricity (that is, the offtaker) is an important factor in assessing the risk associated with a project (that is, if an offtaker is less financially stable, the risk of a default increases, making capital more expensive, which in turn increases the cost of power⁴³). Lastly, we consider whether a PV project being in a solar park has an effect on bid price. Solar parks are designated areas where environmental impact assessment, land procurement and interconnection are already taken care of. However, these increased costs may be reflected in the land price for the solar projects. By differentiating between solar parks and normal land, we are able to capture the price differences between the two approaches.

Thus, we use the following specification to study the effect of LCRs on bid price:

where bid_i is the individual bid of each firm, r is the auction round, LCR_r is the dummy for whether local content was required or not in the auction, Year is the time dummy to control for temporal shocks, Cum_MW is the cumulative installed capacity prior to the given auction in the NSM Phase II, Offtaker is a dummy for the utility buying the power (1 = SECI, 0 = NTPC), Solar park indicates whether the project is within a solar park, Sol refers to the annualized average solar resources (kWh m⁻² d⁻¹) in each state and ε_i is the error term. We also include an interaction term between LCR and our time dummy, to test whether the effect of LCRs changed over time.

Part of the auctioned capacity was tendered under the viability gap funding (VGF) scheme, where the government fixed a base power purchase agreement (PPA) price and companies could request a top-up on the existing base price to make their project financially viable. Since price-only auctions have been implemented in India, the bidders who quoted the lowest amount of VGF were awarded the contracts until the auctioned capacity was reached (it should be noted that bidders were allowed to quote a lower PPA tariff than proposed and waive the VGF, but this rarely happened). Given that the VGF is dispensed as a capacity-based payment at the beginning of the lifetime of a power plant instead of as a constant subsidy for each unit of electricity generated, we had to levelize the amount to compare the outcomes with the generic auction results, where the payments are made across the entire lifetime of the power plant. Therefore, we applied the following method, which is based on the commonly used levelized cost of electricity (LCOE) calculation⁴⁴, to calculate levelized VGF:

In equation (4), E_t is the electricity generated in year t, C is the project’s capacity in MW and Flh is its full-load hours. We assume constant, region-specific full-load hours, which can be found in ref. ⁴⁵. For bids that did not indicate the project’s location in India, we assume a capacity factor of 20% and thus full-load hours of Flh = 1,752 h. Moreover, we assume a discount rate of d = 10% and a plant life of t = 25 years. With our approach, we are also able to capture the time value of money induced by the different VGF disbursement methods applied throughout Phase II (Supplementary Table 7; note that there was no VGF disbursement in Batch II). We then add the resulting levelized VGF support to the specific PPA price.

To estimate the possible range of values for the additional cost borne directly by the Indian government due to LCRs, we used the average estimates from our Heckman regression of the additional cost of power of LCR bids when compared to the open bids. We compute this overall cost to the Indian government via an NPV model in which we discount all future payments from the Indian government to solar power plant owners and compare the cost to the clean technology budget in India. We use discount rates of 10%, 12% and 14% and a capacity factor of 20% for the solar PV plants and a 25-year running time based on REN21 (2018) data. These numbers are roughly similar (apart from possibly lower discount rates in this study) for other developing and emerging economies. These discount rates are based on information used by the Indian government for evaluating public projects⁴⁶. Given the well-known challenges of choosing social discount rates (SDRs)⁴⁷, we perform a sensitivity analysis by varying the discount rate between 10%, 12% and 14%. Taken together, these values for the SDRs encompass typical values of SDRs used in other developing and emerging economies, something that helps make our results more comparable to other countries⁴⁶. We use the average real bid price from all open category auctions as our base price to calculate the additional cost of LCRs over the lifetime of an average solar project subject to LCRs.

In order to analyse the possible benefits of the LCR policy, we select a small set of indicators commonly used in the innovation systems and catching-up literature to determine whether a country is ‘narrowing’ the gap between the innovation leader and itself. While there are no perfect sets of metrics, we employ three different metrics commonly used in the innovation and economics literature: (1) domestic and international patent filings in the technology of interest⁴⁰, (2) domestic production versus international imports⁵ and (3) exports to other countries from the country of interest⁴.

This analysis should be understood as correlational rather than causal, in contrast to the first part of our analysis. In addition, given how relatively recent the policy is, this analysis captures only short-term manufacturing and innovation effects. This is a limitation because some of the impacts of the policy, such as ongoing consolidation of the local industry through mergers and acquisitions, may take more time to materialize. Hence, the main contribution of this paper is the empirical assessment of the additional costs of LCRs, while the analysis of the possible benefits provides indicatory evidence of the evolution of important manufacturing and innovation metrics.

Data

In our analysis, we focus on the NSM Phase II auction results from 2014 to 2017. We did not include the bids and results from NSM Phase I since the majority of the projects (61% of total capacity deployed⁴⁸) in the auction relied on thin film technology (as opposed to silicon panels), which was exempt from LCRs. Furthermore, we focused on the results of the PPA-based scheme and did not consider the EPC programme, which has a different focus: the auctioneer procures and owns the project and does not remunerate the electricity generated over 25 years to the project developer. The different remuneration mechanism, limited availability of data and different auction design elements, as well as different offtakers, hinder the comparability of the data. For the same reason, we neglect auctions conducted by state governments and focus solely on central government tenders conducted by either SECI or NTPC.

Contrary to most other countries conducting auctions, the Indian government shows a high degree of transparency in terms of publishing bids in the NSM Phase II auctions—including information on firms and the bid prices of both successful and unsuccessful applicants. We collected the data about the bid prices and the respective bidders from various government sources, either directly through governmental bodies (for example, SECI) or indirectly through different industry and reputable news sites, such as Mercom India or EQ International Magazine. We include all auction rounds of Phase II that had LCR regulations in place for a total of 28 auction windows across 10 Indian states. As shown in Supplementary Fig. 5, we intentionally excluded from the analysis the state-wise utility-scale PV tenders (around 14 GW by September 2017). In addition, we exclude the central government EPC tenders (1.6 GW) as well as the ‘open category’ rounds in central government auctions in which no counterfactual LCR auction took place (around 4.6 GW), such as the 100 MW auction in Uttar Pradesh in Batch III.

We consider our dataset with 277 bids complete in terms of auction rounds, since LCRs were abolished on 14 December 2017 due to a ruling of the WTO, with NTPC’s 250 MW Indian-wide auction being the last one under LCR regulation (the auction was later cancelled due to the negotiated phase-out of LCR). For further analysis, we consider the available submitted bids, rescale those to 2014 US dollar values to reflect inflation, and use logged bid values in our regression to normalize them. In summary, we consider bids with a total capacity of 21.7 GW, of which 18.7 GW were submitted in the open category and 3 GW under the LCR scheme.

We also collect detailed firm data for all 85 firms within our sample. For each firm, we analyse whether it belongs to a bigger firm. Several firms are so-called special purpose vehicles, which are created merely to bid in a given auction. Given that these firms have access to the human, financial and technical capital of the bigger firm that they belong to, we use the firm characteristics of the parent company. In addition, we collect data on the employment numbers (which could be found for all firms, as opposed to sales numbers, which were unavailable for many privately owned firms), check whether the firm is an SOE and research whether the firms themselves have manufacturing capacities (that is, are vertically integrated). We analyse whether the firm had already bid in the first phase of the NSM, which might give firms a distinct advantage over newcomers due to experience with local regulations. We also check whether the main focus of the company is energy or whether it has just recently diversified its firm activities into energy. Lastly, we analyse whether the firm was founded in India or was registered abroad. We posit that all of these characteristics may influence whether a firm participates in a given auction (for example, we assume that firms that have local manufacturing capabilities are more likely to participate in LCR auctions).

We used solar irradiation maps from the National Renewable Energy Laboratory (NREL) and converted them via QGIS (version 2.8.14) into mean, maximum and minimum values for each state. The NREL dataset provides solar resource in India for surface cells of 0.1 degrees in both latitude and longitude, or nominally 10 km in size. The NREL calculations are based on data from the Meteosat-5 and Meteosat-7 geostationary meteorological satellites⁴².

Patent data were collected from the Indian Patent Database using web scraping methods (Python package Selenium), as the patent office does not offer an application programming interface (API). We employ a typology from a recent, comprehensive review of international patent classification (IPC) terms and their correspondence to PV system components published in Renewable and Sustainable Energy Reviews⁴⁹. This typology covers 284 distinct IPC codes in seven groups: cells, panels, electronics, energy storage, monitoring/testing, devices and combined. Studies comparing global trends in patenting to track innovation normally rely on large patent databases such as the European patent database PATSTAT, which aggregates patent statistics across many domestic offices. However, for India the PATSTAT data are woefully incomplete, leading us to resort to web scraping techniques.

The data on domestic production and imports in Fig. 4c are based on the Directorate General of Trade Remedies investigation on the imposition of safeguards on solar PV cells and modules on behalf of five Indian solar producers. The export data in Fig. 4d were exported from the global United Nations trade database Comtrade using the commodity code ‘HS 854140’, which describes ‘photosensitive semi-conductor devices, including photovoltaic cells whether or not assembled in modules or made up into panels’⁵⁰.

Source: Ecology - nature.com