Narendersingh007/india-solar-benchmark-dataset

India Solar Benchmark Dataset

A large-scale benchmark dataset for solar irradiance forecasting and renewable energy research built from NASA POWER meteorological observations across 50 major Indian cities.

The benchmark contains 10 years of hourly observations (2016–2025) and is distributed as two complementary datasets:

• india_multicity_raw.parquet – cleaned and standardized observations after preprocessing, intended for custom feature engineering and research.
• india_multicity_ml_ready.parquet – fully engineered benchmark containing temporal, geographical, solar geometry, physics-informed, lag, and rolling statistical features.

The benchmark is designed for reproducible research in renewable energy forecasting, time-series machine learning, feature engineering, and physics-informed AI.

---

Links

Resource	Link
GitHub Repository	https://github.com/Narendersingh007/india-solar-benchmark-dataset
Kaggle Dataset	https://www.kaggle.com/datasets/narendersingh007/india-solar-benchmark-dataset

---

Dataset Details

Dataset Statistics

Metric	Value
Cities	50
Years	2016–2025
Frequency	Hourly
Total Records	4,383,600
Features (ML Ready)	60
Raw Source Files	500
Target Variable	ALLSKY_SFC_SW_DWN
Storage Format	Parquet
Source	NASA POWER API

---

Included Files

File	Description
india_multicity_raw.parquet	Cleaned benchmark dataset before feature engineering
india_multicity_ml_ready.parquet	Fully engineered benchmark dataset
train.parquet	Chronological training split
val.parquet	Validation split
test.parquet	Holdout test split
00_sample_preview.csv	Lightweight preview for schema exploration

---

Target Variable

ALLSKY_SFC_SW_DWN

Hourly Global Horizontal Irradiance (GHI) measured in W/m².

This represents the total incoming shortwave solar radiation reaching a horizontal surface and serves as the primary prediction target.

---

Uses

Direct Use

This benchmark is intended for

• Solar irradiance forecasting
• Renewable energy prediction
• Photovoltaic power estimation
• Time-series forecasting research
• Deep learning (LSTM, GRU, Transformer)
• Gradient boosting (XGBoost, LightGBM, CatBoost)
• Feature engineering research
• Physics-informed machine learning
• Renewable energy analytics

---

Out-of-Scope Use

The benchmark should not be used for

• Safety-critical operational decisions
• Grid control without additional validation
• Applications requiring ground-station precision

NASA POWER observations are satellite-derived and modeled products and may differ from local sensor measurements.

---

Dataset Structure

Chronological Split

Split	Years
Train	2016–2022
Validation	2023
Test	2024–2025

Chronological splitting prevents temporal leakage and reflects real-world forecasting workflows.

---

Feature Categories

The ML-ready benchmark contains engineered features grouped into

• Weather variables
• Temporal features
• Cyclical encodings
• Geographic metadata
• Solar geometry (PVLIB)
• Physics-informed variables
• Lag features
• Rolling statistical features

---

Dataset Creation

Motivation

Public benchmark datasets for large-scale solar irradiance forecasting across India are limited.

This benchmark was created to provide a standardized, reproducible dataset for machine learning, renewable energy, and forecasting research.

---

Source Data

Meteorological and solar radiation observations were collected from the NASA POWER API.

Solar geometry and clear-sky irradiance variables were generated using PVLIB.

---

Data Processing Pipeline

The benchmark generation pipeline performs

1. NASA POWER data collection
2. Multi-city aggregation
3. Data cleaning and validation
4. Benchmark raw dataset generation
5. Temporal feature engineering
6. Solar geometry computation
7. Physics-informed feature generation
8. Lag feature creation
9. Rolling statistics computation
10. ML-ready benchmark generation
11. Train / Validation / Test splitting
12. Metadata and validation report generation

---

Bias, Risks and Limitations

Potential limitations include

• NASA POWER modeled observations rather than ground-station measurements
• Coverage limited to selected Indian cities
• Regional climate variability
• Forecast performance depending on downstream modeling choices

Users should independently validate models before operational deployment.

---

Citation

@dataset{india_solar_benchmark_dataset,
  title={India Solar Benchmark Dataset},
  author={Narender Singh},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/Narendersingh007/india-solar-benchmark-dataset}
}

---

Acknowledgements

This benchmark is built using meteorological and solar radiation observations provided by the NASA POWER project.

Solar geometry and clear-sky irradiance variables were generated using PVLIB.

• NASA POWER: https://power.larc.nasa.gov/
• PVLIB: https://pvlib-python.readthedocs.io/

---

Example Usage

import pandas as pd

# ML-ready benchmark
df = pd.read_parquet("india_multicity_ml_ready.parquet")

# Raw benchmark
raw_df = pd.read_parquet("india_multicity_raw.parquet")

print(df.shape)
print(raw_df.shape)