A harmonized inventory of ground-mounted photovoltaic installations in India, 2019–2022

Abstract

# Global PV Inventory — India (2019-2022 backbone, 2022 snapshot) ## Summary This dataset is a deduplicated, attribute-enriched vector inventory of 3,616 ground-mounted, utility-scale photovoltaic (PV) installations in India, produced by fusing three openly-licensed source datasets — the WRI Global Power Plant Database (`wri_gppd`), Kruitwagen et al. (2021)'s global PV inventory (`kruitwagen_2021`), and a global raster-derived PV footprint time series, 2019-2022 (`global_pv_maps_2019_2022`) — through a cross-source spatial conflation and attribute reconciliation pipeline. Each installation record carries a reconciled capacity estimate (57,881.8 MW national total, -7.8% against the IRENA (2023) *Renewable Capacity Statistics 2023* 2022 reference figure of 62,804 MW), a representative footprint or point geometry, per-attribute source provenance, and enrichment fields (land cover class, administrative region, solar irradiance and temperature climatology). 38.9% of installations were confirmed by two or more independent sources; the remainder rely on a single source. This is a **Data Descriptor companion dataset** — see the associated code release and manuscript for the full fusion methodology, validation approach, and known limitations. ## Temporal scope The three backbone sources span different vintages: `global_pv_maps_2019_2022`'s raster detection runs 2019-2022 (per-installation `first_seen_year`, where available, falls in this range); `kruitwagen_2021` and `wri_gppd` are each a single snapshot from their own publication/last-update dates. The reconciled dataset as a whole should be treated as a **2022 snapshot** — the latest extent any backbone source reaches — not a time series. Installations added to the real-world PV fleet after each source's own cutoff are not represented; see `validation_report.json` and the companion manuscript for a discussion of this as a driver of regional under-counts (e.g. Gujarat). ## Spatial reference All geometries are in **EPSG:4326 (WGS 84)**, matching `config.yaml`'s `working_crs` in the companion code release. Installations are ground-mounted, utility-scale PV only (no rooftop/ distributed solar) within India's ADM0 boundary (geoBoundaries), confirmed by a hard polygon clip during enrichment — installations outside India are dropped, not retained with a null region. ## Files | File | Description | |---|---| | `global-pv-inventory.gpkg` | The full inventory as a GeoPackage (single layer), all 25 fields below, EPSG:4326. Recommended format for GIS software (QGIS, ArcGIS). | | `global-pv-inventory.parquet` | The same inventory as GeoParquet. Recommended format for programmatic/columnar analysis (geopandas, DuckDB, etc.). | | `global-pv-inventory.csv` | The same inventory as CSV, with geometry encoded as WKT. Provided for tools without native GeoPackage/GeoParquet support. | | `data_dictionary.json` | Machine-readable field-by-field schema (type, units, definition) for all files above — source for the Fields table below. | | `validation_report.json` | Aggregate-only validation statistics: cross-source agreement rates, national/state capacity totals vs. the IRENA reference, and aggregate OpenStreetMap and TZ-SAM cross-check statistics (counts and sums only — no per-record geometry or attribute data from either source; both are non-redistributable validation-only comparisons, see **Note on validation_report.json** below). | | `figures/agreement_matrix.png` | Cross-source pairwise agreement (Jaccard index) heatmap. | | `figures/capacity_by_region.png` | Choropleth map of reconciled capacity by Indian state/UT. | | `figures/national_total_comparison.png` | Bar chart: this dataset's national total vs. the IRENA 2022 reference. | | `figures/fig_conflation_gallery.png` | 4-panel gallery of real conflation examples (multi-source match, fragment reassembly, point-to-footprint match, panel-vs-site area contrast). | | `figures/fig_match_rule.png` | Schematic diagram of the conflation match rule (candidate generation, IoU/containment confirmation, grouping, fragment merge). | | `figures/fig_provenance.png` | Source-composition (UpSet plot) and capacity-provenance breakdown of the inventory. | **Note on `validation_report.json`**: this file's `osm_crosscheck` and `tzsam_crosscheck` sections report only aggregate counts/sums (e.g. "3,005 of 3,616 installations matched a TZ-SAM asset") — by design, no geometry, attribute, or per-record data from either non-redistributable source is included anywhere in this file or dataset. TZ-SAM (TransitionZero's Solar Asset Mapper) is CC BY-NC 4.0 (non-commercial); OpenStreetMap is ODbL. **A per-state chart derived from TZ-SAM (`tzsam_capacity_comparison.png` in the source project) was deliberately excluded from this release** even though it contains no raw TZ-SAM data, out of caution about attaching any CC BY-NC-derived content to a CC BY 4.0 release — see the code release's `sources_manifest.yaml` for the full source-by-source licensing record. If a fully OSM/TZ-SAM-aggregate-free variant of `validation_report.json` is preferred, one can be produced on request. ## Fields Derived directly from `data_dictionary.json` (25 fields, `global-pv-inventory.gpkg`/`.parquet`/`.csv` share the identical schema). | Field | Type | Units | Definition | |---|---|---|---| | `installation_id` | string | | Unique identifier assigned to each conflated/reconciled PV installation (format PV-NNNNNNN). | | `contributing_sources` | string | comma-separated list | Which backbone source(s) matched into this installation during conflation: wri_gppd, kruitwagen_2021, and/or global_pv_maps_2019_2022. | | `n_contributing_sources` | integer | count | Number of distinct sources contributing to this installation (1-3). | | `agreement_flag` | string | single_source \| multi_source | Whether this installation was found by only one source or matched across multiple sources during conflation. | | `source_record_ids` | string | semicolon-separated list | Raw source_record_id (source:original_id) of every contributing input record, for traceability back to the original ingested data. | | `source_licences` | string | comma-separated list | Licence(s) of the contributing source(s), confirmed fresh from sources_manifest.yaml at export time (not cached from an earlier stage). | | `name` | string | | Installation name, when reported by a contributing source (currently only wri_gppd supplies this). Null otherwise. | | `country_raw` | string | | Raw country code/name as reported by a contributing source (currently only wri_gppd supplies this). Null otherwise. See admin_region for the derived, India-specific alternative. | | `capacity_mw` | float | MW | Reconciled installed capacity, chosen per config.yaml's capacity_priority_sources order (direct-reported capacity > model-estimated capacity > area-derived fallback). See capacity_mw_provenance. | | `capacity_mw_provenance` | string | | Which source (or "area_derived", meaning footprint area x config.yaml's power_density constant) supplied the capacity_mw value. | | `capacity_conflict_flag` | boolean | | True if 2+ contributing sources reported capacity values differing by more than config.yaml's conflict_tolerance_pct. The priority-order value in capacity_mw is still used; this only flags that sources disagreed — see capacity_mw_by_source for the individual values. | | `capacity_mw_by_source` | string | semicolon-separated key=value list | Audit trail of each source's own (within-source-summed) capacity value in MW, e.g. "wri_gppd=12.3000;kruitwagen_2021=15.0000". | | `geometry_type` | string | polygon \| point | Type of the geometry column for this installation. | | `area_km2` | float | km² | True area of the representative geometry, computed in an equal-area projection (config.yaml crs.equal_area_crs). Null for point-only installations. | | `area_km2_provenance` | string | | Which contributing source's polygon was used as the representative geometry/area (the largest single contributing polygon). [†] Null for point-only installations. | | `area_conflict_flag` | boolean | | True if 2+ polygon-contributing sources' own (within-source-summed) footprint areas differ by more than config.yaml's conflict_tolerance_pct. | | `area_km2_by_source` | string | semicolon-separated key=value list | Audit trail of each polygon-contributing source's own summed footprint area in km². | | `first_seen_year` | integer | calendar year | Earliest year any contributing source observed PV at this location (currently only global_pv_maps_2019_2022 supplies this). Null otherwise. | | `first_seen_year_provenance` | string | comma-separated list | Which source(s) supplied the earliest first_seen_year value. | | `land_cover_code` | integer | ESA WorldCover class code | ESA WorldCover class code at the installation's centroid. See land_cover_class for the human-readable label. | | `land_cover_class` | string | | Human-readable ESA WorldCover land-cover class at the installation's centroid (e.g. Built-up, Cropland, Grassland). | | `admin_region` | string | | Indian state/union territory (geoBoundaries ADM1) containing the installation's centroid. [†] Null typically means a rare boundary-precision miss. | | `ALLSKY_SFC_SW_DWN_ann` | float | kWh/m²/day | NASA POWER all-sky surface shortwave downward irradiance, annual climatological mean, nearest available grid point (~1 degree resolution). | | `T2M_ann` | float | degrees Celsius | NASA POWER 2-metre air temperature, annual climatological mean, nearest available grid point (~1 degree resolution). | | `geometry` | geometry (WKT in the CSV export) | EPSG:4326 | Representative footprint (the largest contributing polygon) or point location. | **[†] Known documentation lag**: `data_dictionary.json` (source: a hardcoded dictionary in `pv_fusion/export.py` in the code release) was written before two later pipeline fixes and was not updated to match: - `area_km2_provenance` — the actual current rule (`config.yaml`'s `