arushisinha98/peek-load

Rooftops to Load Curves: building-level electricity demand for Rwanda

6,434,247 building footprints in Rwanda, each with an estimate of what it is used for, how much electricity it would use, when in the day, and how likely it is to be grid-connected. Release v1.0.0 by Peek Load, built for the OSEAS26 hackathon (EnAccess challenge From building footprints to energy demand) with an open, reproducible pipeline: https://github.com/esemsc-as4623/peek-load.

National total (sum of the building table)	GWh/yr
Potential demand (every building grid-connected)	423
Expected demand (households weighted by their probability of a connection, 2024)	320
Expected demand by year (demand_by_year): 2016 back-cast / 2030 projection	205 / 389

Quick start

import pandas as pd
# one district (columns below); data/buildings/ holds the 30 districts + unassigned.parquet
b = pd.read_parquet("hf://datasets/arushisinha98/peek-load/data/buildings/kirehe.parquet")
b.groupby("use").annual_kwh_expected.sum() / 1e6          # GWh/yr by building use

import geopandas as gpd                                     # footprints as polygons (GeoParquet, lon/lat WGS84)
g = gpd.read_parquet("hf://datasets/arushisinha98/peek-load/data/buildings/kirehe.parquet")

import duckdb                                               # the whole country in SQL
duckdb.sql("""SELECT district, round(sum(annual_kwh_expected) / 1e6, 1) AS gwh
              FROM 'hf://datasets/arushisinha98/peek-load/data/buildings/*.parquet' GROUP BY 1 ORDER BY 2 DESC""")

from datasets import load_dataset                           # the Hub loader (configs: buildings, profiles, ...)
profiles = load_dataset("arushisinha98/peek-load", "profiles", split="train")

A building's 24-hour load curve is the sum over its profiles of weight × the profile's per-unit hourly curve:

bp = pd.read_parquet("hf://datasets/arushisinha98/peek-load/data/building_profiles.parquet")
lib = pd.read_parquet("hf://datasets/arushisinha98/peek-load/data/profiles.parquet").set_index("profile_id")
hours = [f"w_h{h:02d}" for h in range(24)]
one = bp[bp.bldg_id == b.bldg_id.iloc[0]]
curve_w = (lib.loc[one.profile_id, hours].to_numpy() * one[["weight"]].to_numpy()).sum(axis=0)  # W, potential
# expected: multiply each row's weight by its p_grid. Area curves: sum building curves first, then take the peak.

Files

Path	Rows	Size	What it holds
data/buildings/<district>.parquet	6,434,247	820 MB	one row per footprint (GeoParquet), one file per district + unassigned
data/benchmarks.parquet	21	7 kB	the model against REG / RURA sales, customers and MoH facility counts
data/building_profiles.parquet	3,523,876	36.0 MB	building x profile: units, energy weight, REG scale, P(connected)
data/demand_by_year.parquet	450	28 kB	district x year, 2016-2030: potential and expected GWh, growth
data/districts.parquet	30	2.7 MB	district totals and boundaries (GeoParquet)
data/profiles.parquet	23	62 kB	RAMP library: per-unit hourly W (all days, weekday, weekend), daily kWh p10/mean/p90
data/sectors.parquet	416	8.3 MB	sector totals and boundaries (GeoParquet)
docs/			datasheet, data sources, status report (decisions), data contract, codebook, benchmarks, AI use

The five challenge outputs

Output	Columns in buildings
1. Estimated peak load	peak_w (expected coincident peak, W), peak_hour (local time, UTC+2)
2. Annual energy consumption	annual_kwh_potential, annual_kwh_expected, annual_kwh_p10, annual_kwh_p90
3. Cooling demand category	cooling_class, tmax_p95_c, cdd22_per_year
4. Load profile characteristics	load_factor, evening_share, daytime_share, night_base_w, weekday_weekend_ratio, load_shape
5. Productive-use viability	pu_viability_score (0-100), pu_viability (low / medium / high)

Columns of buildings

Column	Unit	Meaning
bldg_id		stable id (RWA- + hash of source and geometry); joins every table
province, district, sector, cell, village		administrative areas (geoBoundaries ADM1-5); null outside the boundaries
lon, lat, geometry	degrees	centroid and footprint polygon (WKB, lon/lat on WGS84)
h3_r9, h3_r7		H3 cells (~0.1 km², ~5 km²)
footprint_source, source_confidence, qa_flags		google / microsoft / osm; detector confidence; quality flags (kept, not dropped)
area_m2, height_m, est_floors, gfa_m2, first_seen_year	m², m	footprint area; height (Google 2.5D, 2023); estimated floors; floor area; first year stably detected
facility_type, dist_road_any_m, dist_mv_line_m, pop_density_per_km2, ghsl_class, elevation_m	m, people/km²	context: tagged facility, distance to a road and to a (predicted) MV line, density, settlement class, elevation
setting		urban / rural, matched to the 2022 census share of urban households per district
use		residential, mixed_shop_house, commercial, institutional, religious, productive_use, industrial_warehouse, ancillary (no load)
use_source		tag, facility, census, registry, model or remainder: where the building's use comes from
establishment, establishment_share		profile of the business or institution in it; share when a facility spans several buildings
households, upper_wealth	count	census households placed in the building (up to one per floor); upper wealth group (DHS-calibrated)
use_agreement	0-1	share of 10 perturbed allocations that keep this use: how certain the placement is
annual_kwh_potential, annual_kwh_expected	kWh/yr	if connected; expected = households × P(connected) (establishments count as connected)
annual_kwh_p10, annual_kwh_p90	kWh/yr	day-to-day spread of the profiles (potential)
peak_w, peak_hour, load_factor, evening_share, daytime_share, night_base_w, weekday_weekend_ratio, load_shape	W, h	from the building's 24-hour curve (potential)
cooling_class, tmax_p95_c, cdd22_per_year	°C	climate of the H3 r7 cell (ERA5-Land 2020-2025): low < 26 °C, medium 26-29, high ≥ 29 (tmax p95)
pu_viability_score, pu_viability	0-100	productive-use candidate score (market access, customers, grid, ability to pay, cropland, size)
p_connected_model, p_connected_calibrated	0-1	neighbourhood probability of a grid connection (DHS 2025 model); raked to EICV7 2024 per district
expected_kwh_2016 ... expected_kwh_2023	kWh/yr	back-cast expected demand (presence-weighted, so it can be fractional)

How the numbers are made

1. Footprints and context. VIDA's merge of Google, Microsoft and OpenStreetMap footprints; heights and yearly presence (Google Open Buildings 2.5D Temporal); roads, places, settlement class, population density, climate (Copernicus ERA5-Land) and night lights (NASA Black Marble).
2. Use. A gradient-boosting classifier gives every building use probabilities. Census and registry counts are then placed on the most likely buildings, rarest use first: Establishment Census 2023 (manufacturing, services, shops), MINEDUC schools, Ministry of Health facilities (63 hospitals, 518 health centres, ...), places of worship, and the 2022 census households per district (urban and rural). The rest are ancillary structures.
3. Load profiles. 23 RAMP appliance profiles (households by setting and wealth, shops, bars, salons, mills, workshops, schools, health facilities, offices, churches, stores) built from World Bank MTF 2022 appliance use, a cited evidence base and SEforALL's 2023 health-facility measurements. Household energy is scaled to REG's mean residential consumption (Kigali / elsewhere).
4. Connection. A model trained on DHS 2025 gives each neighbourhood a probability of grid connection, raked to EICV7 2024 per district.
5. Years. Back-cast 2016-2023 from each building's detected presence, pinned to census household growth, with connection rolled back per district to EICV4 and EICV5; projected to 2030 by district growth.

Checked against REG (data/benchmarks.parquet, docs/benchmarks.md): households' expected demand is 208 GWh in 2023 against 211 GWh of residential sales; connected households per REG branch correlate at r = 0.93; health facilities use 23.0 GWh against REG's 22.7. Businesses are under-counted (about 30% of REG's non-residential sales), and large hotels are missing.

How to read a building's values. Most inputs are aggregates (district counts, survey clusters, national means). Building values are a plausible allocation that reproduces those totals, not observations: use use_agreement to see how stable a building's use is, and prefer area sums for planning.

Limitations

Building use is uncertain where tags are absent; heights are compressed (floors above 2 are unreliable); rural manufacturing is 25% electric mills by assumption; service splits are assumptions; private clinics use a health centre's profile; profiles assume a grid connection; seasonal variation is not modelled; businesses are under-counted. More in docs/datasheet.md and docs/status_report.md.

Licence and attribution

Open Database License (ODbL-1.0), see LICENSE.md. Footprints: VIDA combined Open Buildings (Google, Microsoft, OpenStreetMap), ODbL-1.0. Climate: Copernicus ERA5-Land (Copernicus licence, attribution). Night lights: NASA Black Marble (open). Survey-derived values (DHS 2025, World Bank MTF 2022, NISR EICV) enter only as model outputs and aggregates; no microdata, no DHS cluster locations. Meta's Relative Wealth Index values are not included. Please cite this dataset and the underlying sources (docs/data_sources.md).

@misc{peek_load_2026_rooftops,
  author       = {Peek Load},
  title        = {Rooftops to Load Curves: building-level electricity demand for Rwanda},
  year         = {2026},
  version      = {v1.0.0},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/arushisinha98/peek-load}},
  note         = {OSEAS26 hackathon, EnAccess challenge. Code: https://github.com/esemsc-as4623/peek-load}
}