This dataset is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence (CC BY-NC-SA 4.0): research, academic, personal and other non-commercial use is permitted with attribution; adaptations must be shared under the same licence; commercial use is not permitted. The original public source data remains under the Government Open Data License - India (GODL-India). Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.

BharatEnergy-Bench

Licence: CC BY-NC-SA 4.0 — research and non-commercial use only; derivatives must be shared under the same licence. Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai

Access is approved automatically when the email you enter matches your Hugging Face account email and belongs to the organisation you name; other requests are reviewed manually.

A probabilistic forecasting benchmark for India's electricity supply: the actual daily generation of every power station and every thermal and nuclear generating unit in the Central Electricity Authority's daily generation report. Both tracks hold out everything from 2025-07-01 on as the test period and forecast every series from rolling monthly origins, 1, 7 and 30 days ahead, with nine quantiles. The release ships 1,067 cleaned series, the task definitions, the forecasts of 7 models and an evaluate.py that recomputes the leaderboard from those forecasts.

At 1 day, the lowest overall WQL is BharatEnergy-Chronos (bharatenergy_chronos) (0.0897); 2.0% below chronos2, 6.1% below naive.
At 7 days, the lowest overall WQL is BharatEnergy-Chronos (bharatenergy_chronos) (0.1381); 3.2% below chronos2, 12.4% below lightgbm.
At 30 days, the lowest overall WQL is BharatEnergy-Chronos (bharatenergy_chronos) (0.1757); 5.3% below chronos2, 15.3% below lightgbm.
BharatEnergy-Chronos has the lowest WQL in 5 of 6 track x horizon tasks.

Coverage: 459 power stations and 608 generating units across 33 states/UTs and 5 regions (+ Bhutan imports), daily actual generation 2022-01-01 to 2026-01-02

Headline results

Overall = mean of the two daily tracks (station_generation, unit_generation), each averaged over its test origins. Lower WQL and MASE are better; a calibrated 80% interval covers 0.80.

h (days)	model	WQL	MASE	cov_80
1	BharatEnergy-Chronos (`bharatenergy_chronos`)	0.0897	0.5900	0.7380
1	`chronos2`	0.0915	0.6070	0.7413
1	`timesfm25`	0.0919	0.6060	0.7831
1	`naive`	0.0956	0.5955	0.7469
1	`lightgbm`	0.1128	0.6842	0.7626
1	`seasonal_naive`	0.1845	1.1271	0.7961
1	`climatology`	0.2271	1.7031	0.7151
7	BharatEnergy-Chronos (`bharatenergy_chronos`)	0.1381	0.8812	0.7593
7	`chronos2`	0.1427	0.9239	0.7560
7	`timesfm25`	0.1453	0.9360	0.7815
7	`lightgbm`	0.1576	0.9659	0.7413
7	`naive`	0.1660	0.9980	0.7606
7	`seasonal_naive`	0.1933	1.1575	0.7784
7	`climatology`	0.2260	1.7086	0.7219
30	BharatEnergy-Chronos (`bharatenergy_chronos`)	0.1757	1.2044	0.7671
30	`chronos2`	0.1855	1.3140	0.7513
30	`timesfm25`	0.1903	1.3288	0.7800
30	`lightgbm`	0.2073	1.3566	0.7233
30	`climatology`	0.2320	1.7600	0.6823
30	`naive`	0.2347	1.4803	0.7802
30	`seasonal_naive`	0.2405	1.5184	0.8079

Per track (WQL and MASE by horizon):

Power-station daily generation (`station_generation`)

model	WQL h=1	WQL h=7	WQL h=30	MASE h=1	MASE h=7	MASE h=30
BharatEnergy-Chronos (`bharatenergy_chronos`)	0.0855	0.1261	0.1594	0.6554	0.9298	1.2578
`chronos2`	0.0855	0.1287	0.1656	0.6637	0.9742	1.3816
`timesfm25`	0.0853	0.1301	0.1693	0.6541	0.9802	1.3956
`lightgbm`	0.1001	0.1377	0.1876	0.7802	1.0545	1.4751
`naive`	0.0894	0.1490	0.2115	0.6566	1.0542	1.5568
`seasonal_naive`	0.1648	0.1697	0.2123	1.2246	1.2148	1.5966
`climatology`	0.2071	0.2048	0.2134	1.8733	1.8631	1.9328

Generating-unit daily generation (`unit_generation`)

model	WQL h=1	WQL h=7	WQL h=30	MASE h=1	MASE h=7	MASE h=30
BharatEnergy-Chronos (`bharatenergy_chronos`)	0.0940	0.1501	0.1919	0.5246	0.8326	1.1510
`chronos2`	0.0976	0.1567	0.2055	0.5504	0.8737	1.2463
`timesfm25`	0.0985	0.1605	0.2112	0.5579	0.8917	1.2619
`lightgbm`	0.1255	0.1776	0.2271	0.5882	0.8774	1.2380
`climatology`	0.2472	0.2471	0.2506	1.5328	1.5540	1.5871
`naive`	0.1018	0.1830	0.2580	0.5344	0.9419	1.4038
`seasonal_naive`	0.2041	0.2168	0.2688	1.0297	1.1002	1.4402

Power-station daily generation by fuel type (WQL, lower is better; best per horizon in bold)

fuel	windows	BharatEnergy-Chronos (`bharatenergy_chronos`) WQL h=7	`chronos2` WQL h=7	`timesfm25` WQL h=7	`lightgbm` WQL h=7	`seasonal_naive` WQL h=7	BharatEnergy-Chronos (`bharatenergy_chronos`) WQL h=30	`chronos2` WQL h=30	`timesfm25` WQL h=30	`lightgbm` WQL h=30	`seasonal_naive` WQL h=30
HYDRO	1,268	0.1490	0.1573	0.1634	0.1742	0.2150	0.2042	0.2458	0.2580	0.2788	0.3333
NUCLEAR	42	0.0923	0.0926	0.0957	0.1005	0.1114	0.1440	0.1298	0.1405	0.1324	0.1567
THER (DG)	6	0.0081	0.0086	0.0086	0.0156	0.0113	0.0100	0.0109	0.0104	0.0258	0.0128
THER (GT)	230	0.2267	0.2332	0.2510	0.2989	0.3704	0.3314	0.3307	0.3460	0.4133	0.4651
THERMAL	1,131	0.1229	0.1245	0.1250	0.1311	0.1620	0.1512	0.1530	0.1554	0.1733	0.1941

Power-station daily generation by grid region (WQL, lower is better; best per horizon in bold)

region	windows	BharatEnergy-Chronos (`bharatenergy_chronos`) WQL h=7	`chronos2` WQL h=7	`timesfm25` WQL h=7	`lightgbm` WQL h=7	`seasonal_naive` WQL h=7	BharatEnergy-Chronos (`bharatenergy_chronos`) WQL h=30	`chronos2` WQL h=30	`timesfm25` WQL h=30	`lightgbm` WQL h=30	`seasonal_naive` WQL h=30
EASTERN	336	0.1008	0.1006	0.1009	0.1078	0.1387	0.1191	0.1249	0.1267	0.1389	0.1614
NORTH EASTERN	141	0.1019	0.1036	0.1067	0.1174	0.1332	0.1327	0.1440	0.1492	0.1609	0.1769
NORTHERN	807	0.1328	0.1350	0.1373	0.1455	0.1760	0.1621	0.1733	0.1749	0.1964	0.2303
SOUTHERN	715	0.1466	0.1540	0.1555	0.1652	0.2084	0.1979	0.2093	0.2165	0.2357	0.2641
WESTERN	678	0.1244	0.1263	0.1275	0.1331	0.1611	0.1582	0.1580	0.1625	0.1811	0.1988

Generating-unit daily generation by fuel type (WQL, lower is better; best per horizon in bold)

fuel	windows	BharatEnergy-Chronos (`bharatenergy_chronos`) WQL h=7	`chronos2` WQL h=7	`timesfm25` WQL h=7	`lightgbm` WQL h=7	`seasonal_naive` WQL h=7	BharatEnergy-Chronos (`bharatenergy_chronos`) WQL h=30	`chronos2` WQL h=30	`timesfm25` WQL h=30	`lightgbm` WQL h=30	`seasonal_naive` WQL h=30
NUCLEAR	126	0.0971	0.1031	0.1002	0.1180	0.1252	0.1808	0.1687	0.1756	0.1841	0.1948
THERMAL	3,488	0.1526	0.1592	0.1634	0.1803	0.2210	0.1925	0.2070	0.2126	0.2289	0.2719

Generating-unit daily generation by grid region (WQL, lower is better; best per horizon in bold)

region	windows	BharatEnergy-Chronos (`bharatenergy_chronos`) WQL h=7	`chronos2` WQL h=7	`timesfm25` WQL h=7	`lightgbm` WQL h=7	`seasonal_naive` WQL h=7	BharatEnergy-Chronos (`bharatenergy_chronos`) WQL h=30	`chronos2` WQL h=30	`timesfm25` WQL h=30	`lightgbm` WQL h=30	`seasonal_naive` WQL h=30
EASTERN	612	0.1149	0.1180	0.1220	0.1355	0.1754	0.1474	0.1607	0.1648	0.1787	0.2165
NORTH EASTERN	18	0.0843	0.0844	0.0790	0.0919	0.1017	0.0759	0.0783	0.0728	0.0829	0.0974
NORTHERN	918	0.1530	0.1606	0.1644	0.1890	0.2203	0.1919	0.2100	0.2126	0.2362	0.2775
SOUTHERN	755	0.1852	0.1970	0.2028	0.2319	0.2850	0.2497	0.2652	0.2782	0.3016	0.3507
WESTERN	1,311	0.1503	0.1557	0.1586	0.1677	0.2041	0.1884	0.1987	0.2036	0.2125	0.2533

The full table with 80% coverage and window counts is in leaderboard.md / leaderboard.json and in the BharatEnergy-Bench-Leaderboard Space.

Tasks

track	what	unit	freq	series	observations	span	test cutoff	origins	horizons
`station_generation`	actual daily generation of each power station (station total) in the CEA report	MU/day (million kWh)	daily	459	619,832	2022-01-01 to 2026-01-02	2025-07-01	6 (2025-07-01 to 2025-12-01)	1, 7, 30
`unit_generation`	actual daily generation of each thermal and nuclear generating unit	MU/day (million kWh)	daily	608	856,142	2022-01-01 to 2026-01-02	2025-07-01	6 (2025-07-01 to 2025-12-01)	1, 7, 30

Training data for any model = observations strictly before the track's test cutoff; every origin lies on or after the cutoff. A window needs at least 8 observed steps in the context and one observed step in the horizon. Models get the 512 days before the origin as context.

Quickstart

import pandas as pd
from huggingface_hub import snapshot_download

root = snapshot_download("judicialmind/BharatEnergy-Bench", repo_type="dataset")
s = pd.read_parquet(f"{root}/data/station_generation/series.parquet")
print(s.head())

import sys; sys.path.insert(0, root)
from evaluate import evaluate
lb = evaluate(root, models=["seasonal_naive", "chronos2"], only_tracks=["station_generation"])
print(lb.round(4))

To score your own model, write forecasts/<track>/<your_model>/<origin>.parquet (columns series_id, origin, h, q0.1 ... q0.9) for the origins in data/<track>/splits.json and run python evaluate.py --models <your_model>. python evaluate.py --check recomputes the shipped leaderboard; python evaluate.py --by fuel (or --by region) scores per fuel type or grid region.

Files

path	content
`data/<track>/series.parquet`	long series table: `series_id`, `ts`, `target` and static columns
`data/<track>/splits.json`	cutoff, origins, horizons, frequency, seasonality
`data/<track>/tasks.json`	one task per origin x horizon (fev-style)
`windows/<track>.parquet`	scored windows (`series_id`, `origin`) with their MASE scale
`forecasts/<track>/<model>/<origin>.parquet`	quantile forecasts of every model (84 files)
`leaderboard.json`, `leaderboard.md`	scores per track x model x horizon, and the overall table
`leaderboard_breakdown.json`	scores per track x model x horizon x fuel type / grid region
`evaluate.py`, `requirements.txt`	scoring code (numpy, pandas, pyarrow)

Metrics

WQL: mean over the quantiles 0.1-0.9 of 2 x summed pinball loss / summed |y|, pooled over the windows of an origin x horizon task, then averaged over origins.
MASE: mean absolute error of the median divided by the in-context mean absolute seasonal difference (mase_season of splits.json; lag 1 when no seasonal pair exists), averaged over windows and origins.
cov_80: share of observed values inside the 0.1-0.9 interval.

Models

model	description
`bharatenergy_chronos`	BharatEnergy-Chronos: Chronos-2 fine-tuned on train-period generation series
`chronos2`	amazon/chronos-2, zero-shot
`timesfm25`	google/timesfm-2.5-200m, zero-shot
`lightgbm`	global LightGBM quantile model on lagged values
`climatology`	same calendar window in earlier years
`seasonal_naive`	same weekday last week
`naive`	last observed value

Source data

Derived from the Central Electricity Authority (CEA) daily generation report (region-, state-, sector-, type-, station- and unit-wise actual and programmed generation), published on NITI Aayog's National Data and Analytics Platform (NDAP) and the Open Government Data Platform India (data.gov.in) under the Government Open Data License - India (GODL-India). Cleaning: duplicate (series, day) rows averaged, values outside physical bounds (negative generation, or more than 1.1 x monitored capacity x 24 h) flagged and left out of the series, gaps left missing, series with less than a year of history before the cutoff or under 60 test-period days dropped.

Original data: published by Central Electricity Authority (CEA), Ministry of Power on the Open Government Data Platform India (data.gov.in) and, where marked, NITI Aayog's National Data and Analytics Platform (NDAP), under the Government Open Data License - India (GODL-India). The original public data remains under GODL-India; the CC BY-NC-SA 4.0 licence of this repository covers JudicialMind's compilation, derived fields, annotations, benchmarks and models.

Citation

@misc{judicialmind2026bharatenergybench,
  title={BharatEnergy-Bench: Forecasting Daily Electricity Generation of India's Power Stations},
  author={JudicialMind},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/judicialmind/BharatEnergy-Bench}}
}

License

CC BY-NC-SA 4.0 (see LICENSE): research, academic, personal and other non-commercial use, with attribution to JudicialMind and to the original data publishers; adaptations must be shared under the same licence. Commercial use requires a separate licence.

Commercial use: contact surya@judicialmind.ai or hello@judicialmind.ai.