Real-world multivariate time-series dataset collected from the Karnataka State Load Dispatch Centre (SLDC), India (April 22 – August 14, 2024). Combines 15-minute grid-operational telemetry with co-located OpenWeatherMap meteorological observations. Designed for short-term load forecasting (STLF) research with four simultaneous prediction targets: grid frequency, state demand, state generation, and CGS exchange.

Dataset Details

Dataset Description

Karnataka manages electricity dispatch for ~64 million residents across six distribution companies (DISCOMs). The SLDC publishes grid frequency, state demand/generation, DISCOM-wise consumption, and Central Government Scheme (CGS) net exchange figures at ~15-minute intervals. This dataset merges that telemetry with weather observations and provides ready-to-use train/test/validation splits and 5-fold time-series cross-validation folds.

Curated by: Tapan Kumar Das, IIIT Delhi
Language(s): English
License: CC BY 4.0
Collection period: 2024-04-22 to 2024-08-14
Sampling interval: ~15 minutes
Total rows (preprocessed): 8,034
Input features (X): 18 (7 grid + 10 meteorological + 1 categorical)
Target variables (Y): 4 (current_frequency, state_demand, state_generation, cgs)

Dataset Sources

Repository: https://huggingface.co/datasets/happyman/Karnataka-SLDC
Paper: Short-Term Load Forecasting for Regional Smart Grid Operations: A Hierarchical Multi-Task Deep Learning Approach Using the Karnataka-SLDC Dataset (Journal of Supercomputing, under review, 2024)
Grid data source: Karnataka SLDC real-time portal (public)
Weather data source: OpenWeatherMap Current Weather API (Bangalore, 12.9716°N 77.5946°E)

Uses

Direct Use

Short-term load forecasting benchmarks for Indian regional grids
Multi-target / multi-task time-series regression
Reproducing the H-MTL-GRU paper experiments (use paper_split config with scaled CSVs)
Evaluating new forecasting architectures (use general_split config with 5-fold CV)
Grid frequency stability analysis

Out-of-Scope Use

Long-term (seasonal / annual) load forecasting — the dataset covers only 116 days
Generalisation to other Indian states without retraining — Karnataka grid characteristics may differ
Real-time dispatch decision-making without domain expert oversight

Dataset Structure

Data Fields

Input Features — X (18 columns + TIME STAMP)

Column	Unit	Description
`ncep_generation`	MW	NCEP / NTPC-Karnataka scheduled power
`bescom`	MW	BESCOM zone load
`hescom`	MW	HESCOM zone load
`cesc`	MW	CESC zone load
`gescom`	MW	GESCOM zone load
`mescom`	MW	MESCOM zone load
`pavagada_kspdcl`	MW	Pavagada solar park output (KSPDCL)
`weather_temp`	°C	Ambient temperature
`weather_feels_like`	°C	Apparent temperature
`weather_temp_min`	°C	Hourly minimum temperature
`weather_temp_max`	°C	Hourly maximum temperature
`weather_temp_pressure`	hPa	Atmospheric pressure
`weather_temp_humidity`	%	Relative humidity
`weather_temp_visibility`	m	Visibility (max 10,000 m)
`weather_wind_speed`	m/s	Wind speed
`weather_wind_deg`	°	Wind direction
`weather_clouds_all`	%	Cloud cover
`weather_description`	—	Sky condition string (e.g. "light rain")

Target Variables — Y (4 columns + TIME STAMP)

Column	Unit	Description
`current_frequency`	Hz	Grid frequency (nominal 50.0 Hz)
`state_demand`	MW	Karnataka total electricity demand
`state_generation`	MW	Karnataka total in-state generation
`cgs`	MW	CGS net exchange (negative = net export)

Data Splits

Paper Split (80 / 20 chronological)

Partition	Rows	Date Range
Train	6,427	2024-04-22 → 2024-07-30
Test	1,607	2024-07-30 → 2024-08-14

General Split (70 / 10 / 20 chronological)

Partition	Rows	Date Range
Train	5,623	2024-04-22 → 2024-07-24
Val	803	2024-07-24 → 2024-07-30
Test	1,608	2024-07-30 → 2024-08-14

5-Fold Time-Series Cross-Validation (TimeSeriesSplit, expanding window)

Fold	Train rows	Val rows	Val end date
1	1,339	1,339	2024-06-27
2	2,678	1,339	2024-07-04
3	4,017	1,339	2024-07-22
4	5,356	1,339	2024-08-02
5	6,695	1,339	2024-08-14

Full per-fold date ranges are in data/kfold/fold_metadata.json.

Normalization (Paper Protocol)

*_scaled.csv files use leakage-free scaling fitted on training data only:

X numerical (17 cols): MinMax [0, 1]
X categorical: weather_description → one-hot encoded (prefix wd_, ~9 binary cols)
Y current_frequency: shift by −49.0 Hz
Y state_demand, state_generation, cgs: MinMax [0, 1]

Dataset Creation

Curation Rationale

No public benchmark exists for Indian state-level SLDC data covering simultaneous frequency, demand, generation, and CGS targets. This dataset fills that gap and enables reproducible evaluation of multi-task deep learning forecasters at the Karnataka SLDC scale.

Source Data

Data Collection and Processing

Grid telemetry was scraped from the Karnataka SLDC public real-time portal at ~15-minute resolution from April 22 to August 14, 2024. Weather observations were retrieved from the OpenWeatherMap Current Weather API at the Bangalore city centroid (12.9716°N, 77.5946°E) at matching timestamps.

The two sources were merged on exact TIME STAMP. Four administrative columns (current_time, filled_at, weather_temp_sunrise, weather_temp_sunset) were dropped. Rows with any missing value were removed (4 rows), yielding 8,034 clean records.

Who are the source data producers?

Grid data: Karnataka State Load Dispatch Centre (KSLDC), Government of Karnataka, India

Personal and Sensitive Information

This dataset contains no personal or individually identifiable information. All data describe aggregate grid and environmental measurements at the state level.

Bias, Risks, and Limitations

Short temporal window (116 days): Only covers April–August 2024. Winter demand peaks, festival-season spikes, and annual seasonal cycles are absent.
Single geographic region: Karnataka state only. Models trained here may not generalise to other Indian SLDCs.
Single weather station: Bangalore co-ordinates used for all weather features. Karnataka spans arid, semi-arid, and coastal zones — sub-regional weather variation is not captured.
No holiday or event annotations: Public holidays, industrial shutdowns, and IPL/cricket events cause demand anomalies that appear as unexplained noise.
CGS sign convention: Negative CGS values indicate net export (common during Pavagada solar midday surplus). Users should be aware of this accounting convention before modelling.
Data scraper gaps: A small number of 15-minute slots may be missing due to portal downtime; these appear as absent rows rather than NaN-filled rows.

Recommendations

For annual or seasonal forecasting tasks, augment with additional years of SLDC data.
For multi-zone analysis, collect DISCOM-level weather observations rather than a single city coordinate.
When comparing results with other papers, confirm whether the 80/20 paper split or the 70/10/20 general split was used.

Usage

# Load directly from the hub — no download required
import pandas as pd

BASE = "https://huggingface.co/datasets/happyman/Karnataka-SLDC/resolve/main"

# Preprocessed full dataset
df = pd.read_csv(f"{BASE}/data/preprocessed/Karnataka_SLDC.csv", parse_dates=["TIME STAMP"])

# Paper split — scaled, ready for deep learning
X_train = pd.read_csv(f"{BASE}/data/paper_split/X_train_scaled.csv", parse_dates=["TIME STAMP"])
X_test  = pd.read_csv(f"{BASE}/data/paper_split/X_test_scaled.csv",  parse_dates=["TIME STAMP"])
Y_train = pd.read_csv(f"{BASE}/data/paper_split/Y_train_scaled.csv", parse_dates=["TIME STAMP"])
Y_test  = pd.read_csv(f"{BASE}/data/paper_split/Y_test_scaled.csv",  parse_dates=["TIME STAMP"])

print(f"Train: {len(X_train)} rows × {len(X_train.columns)-1} features")
print(f"Test : {len(X_test)} rows")

# Load as HuggingFace Dataset object
from datasets import load_dataset

ds = load_dataset(
    "happyman/Karnataka-SLDC",
    "paper_split_scaled",   # config name
)
df_train = ds["train"].to_pandas()

# 5-fold cross-validation loop
import json, requests, pandas as pd

BASE = "https://huggingface.co/datasets/happyman/Karnataka-SLDC/resolve/main"
meta = requests.get(f"{BASE}/data/kfold/fold_metadata.json").json()

for fold in meta:
    k    = fold["fold"]
    X_tr = pd.read_csv(f"{BASE}/data/kfold/fold_{k}/X_train.csv", parse_dates=["TIME STAMP"])
    Y_tr = pd.read_csv(f"{BASE}/data/kfold/fold_{k}/Y_train.csv", parse_dates=["TIME STAMP"])
    X_va = pd.read_csv(f"{BASE}/data/kfold/fold_{k}/X_val.csv",   parse_dates=["TIME STAMP"])
    Y_va = pd.read_csv(f"{BASE}/data/kfold/fold_{k}/Y_val.csv",   parse_dates=["TIME STAMP"])
    print(f"Fold {k}: train={fold['train_rows']:,}  val={fold['val_rows']:,}")
    # fit your model on X_tr / Y_tr, evaluate on X_va / Y_va

# Create lag-window sequences for PyTorch (paper lag = 17)
import numpy as np, pandas as pd, torch
from torch.utils.data import TensorDataset, DataLoader

BASE = "https://huggingface.co/datasets/happyman/Karnataka-SLDC/resolve/main"
LAG  = 17

X = pd.read_csv(f"{BASE}/data/paper_split/X_train_scaled.csv").drop(columns=["TIME STAMP"]).values
Y = pd.read_csv(f"{BASE}/data/paper_split/Y_train_scaled.csv").drop(columns=["TIME STAMP"]).values

def make_windows(X, Y, lag):
    xs = [X[i-lag:i] for i in range(lag, len(X))]
    ys = [Y[i]        for i in range(lag, len(Y))]
    return np.array(xs, dtype=np.float32), np.array(ys, dtype=np.float32)

X_seq, Y_seq = make_windows(X, Y, LAG)
print(f"Sequences: {X_seq.shape}")   # (6410, 17, 25)

loader = DataLoader(
    TensorDataset(torch.from_numpy(X_seq), torch.from_numpy(Y_seq)),
    batch_size=64, shuffle=False,
)

Citation

@article{das2024karnataka_stlf,
  title   = {Short-Term Load Forecasting for Regional Smart Grid Operations:
             A Hierarchical Multi-Task Deep Learning Approach
             Using the Karnataka-SLDC Dataset},
  author  = {Das, Tapan Kumar and others},
  journal = {The Journal of Supercomputing},
  year    = {2024},
  note    = {Under review}
}

Dataset Card Contact

Contact the corresponding author via the Journal of Supercomputing submission system.

Downloads last month: 14

Total file size:

15.2 MB