Solar resource and the cost of capital across seventeen Indonesian sites: analysis pipeline, logs and derived tables

Abstract

Five complete years of hourly photovoltaic output, 2019 to 2023, retrieved from PVGIS (radiation database PVGIS-ERA5) for seventeen Indonesian sites, together with the analysis that reduces them to capacity factors, specific yields, interannual spreads and levelised costs of energy across a range of costs of capital. The archive supports Chapter 3 of the author's doctoral dissertation, "Artificial Intelligence for Low-Carbon Energy in the Global South: Assessing Impacts and Pathways in Indonesia" (Tsinghua University). It was assembled to test one claim. Indonesia's national electricity plan, the RUPTL 2025-2034, designates Sumba an Iconic Renewable Energy Island and asserts that the island holds the country's best solar resource. It publishes no supporting table. The claim survives the test. Waingapu, in East Sumba, ranks first of the seventeen sites at a five-year mean capacity factor of 19.10 per cent and a specific yield of 1,673 kWh per kWp per year, and it leads its closest national rival in each of the five years taken separately (paired t = 7.87, five years of five). The larger result is an asymmetry. Holding the cost of capital fixed at 6 per cent real, the gap between the best and the worst of the four Sumba load centres is 4.5 USD per MWh. Holding the site fixed at Waingapu and moving across the plausible range of real costs of capital, from 4 to 12 per cent, the levelised cost moves by 42.3 USD per MWh. Financing terms move the answer some 9.5 times further than the choice of site does. A secondary finding concerns the reliability of an imported computational tool. PVGIS's own tilt and azimuth optimiser returns physically impossible configurations at three of the seventeen sites and no answer at all at a fourth, while the textbook equator-facing fixed configuration succeeds everywhere and, at Waingapu, outperforms the optimiser's own recommendation by 1.3 per cent. Both configurations are included so that the comparison can be checked. Contents: the two analysis scripts; six derived tables covering both configurations; the retrieval logs, recording every request made and every point PVGIS refused; the resource-bias sensitivity run of 26 September 2026 with its log; two figures; and every PVGIS response the analysis used, 33 files of 43,824 hourly records each, so that every published number can be re-derived without contacting the service. Levelised costs assume 1,100 USD per kW installed, operation and maintenance at 1.5 per cent of capital a year, a 25-year life and 0.5 per cent annual degradation. The installed cost is an assumption rather than a measurement, and the levelised cost is exactly proportional to it, so a reader preferring a different figure can rescale every cost here by the ratio of the two. No non-public data of any kind is used. Code is released under the MIT licence; data files and retrieved responses under CC BY 4.0. README.txt describes every file, how to reproduce the results, and four things a reader should know before using them. A companion archive, supporting a different chapter of the same dissertation, is deposited separately: "AI for energy: a bibliometric count of application areas, 2015-2025", doi:10.5281/zenodo.22958665. The two archives share no data and neither depends on the other.