Readme.md

GHI Prediction — Solar Radiation Forecasting System

This interactive demo presents a production-grade machine learning system for predicting Global Horizontal Irradiance (GHI) using real-world weather data collected from weather stations across Saudi Arabia.

The model enables fast and accurate solar radiation forecasting, supporting applications in solar energy planning, grid optimization, and sustainable infrastructure development.

---

Input Features & Valid Ranges

All inputs are validated using real data bounds to ensure reliable predictions:

Feature	Valid Range
Air Temperature (°C)	8.10 – 39.70
Air Temperature Uncertainty (°C)	0.50 – 0.60
Wind Direction (°N)	0 – 359
Wind Direction Uncertainty (°N)	0.0 – 4.80
Wind Speed at 3m (m/s)	0.0 – 7.80
Wind Speed Uncertainty (m/s)	0.0 – 0.10
Wind Speed Std Dev (m/s)	0.0 – 4.50
DHI (Wh/m²)	865.8 – 6394.6
DHI Uncertainty (Wh/m²)	37.3 – 1344.8
Standard Deviation DHI (Wh/m²)	148.3 – 1649.6
DNI (Wh/m²)	0.3 – 9517.0
DNI Uncertainty (Wh/m²)	0.2 – 2115.8
Standard Deviation DNI (Wh/m²)	0.4 – 3282.2
GHI Uncertainty (Wh/m²)	98.4 – 5249.6
Standard Deviation GHI (Wh/m²)	117.6 – 1720.7
Peak Wind Speed at 3m (m/s)	0.0 – 40.5
Peak Wind Speed Uncertainty (m/s)	0.0 – 0.20
Relative Humidity (%)	8.3 – 95.9
Relative Humidity Uncertainty (%)	2.8 – 3.7
Barometric Pressure (hPa)	831.3 – 1019.7
Barometric Pressure Uncertainty (hPa)	4.1 – 6.2

---

Model & Training Summary

Multiple machine learning models were trained and evaluated:

• Linear Regression (Deployed Model)
• Random Forest
• Decision Tree
• KNN
• SVR
• XGBoost
• Artificial Neural Network
• Histogram Gradient Boosting

Final Model Selection Strategy

The final deployed model — Linear Regression — was selected using an automated ranking system that balances:

• Prediction accuracy (RMSE, R²)
• Inference latency (real-time performance)

This approach ensures both high accuracy and fast response for real-world deployment.

Best Model Performance

Model	MAE	RMSE	R²	Training Time
Linear Regression	135.94	193.74	0.97	0.01s
Histogram Gradient Boosting	140.58	192.11	0.98	0.83s
XGBoost	148.37	198.05	0.97	2.73s
ANN	154.4	241.22	0.96	10.6s

---

Experiment Tracking

All experiments, comparisons, and model selection were tracked using Weights & Biases with full reproducibility and visualization.

🔗 W&B Dashboard:
https://wandb.ai/nishnarudkar-d-y-patil-university/solar-radiation-prediction

---

Full Project Ecosystem

• 🖥️ Live Web Application:
  https://solar-radiation-prediction-using-saudi.onrender.com
• 💻 Source Code (GitHub):
  https://github.com/nishnarudkar/Solar-Radiation-Prediction-using-Saudi-Arabia-Dataset
• 🧑‍🔬 AWS Builder Center Article:
  https://builder.aws.com/content/36qXvwTipuNwbZMUPf3hQ7n2IS0/from-experiment-tracking-to-automated-model-selection-a-solar-radiation-prediction-pipeline
• 📰 Medium Article:
  https://medium.com/@nishnarudkar/from-experiment-tracking-to-automated-model-selection-a-practical-ml-workflow-cf193d1b1098

---

Team

• Nishant Narudkar
• Maitreya Pawar
• Vatsal Parmar
• Aamir Sarang

---

This Hugging Face Space demonstrates the final trained model with validated real-world constraints and a clean interactive interface for public experimentation.