Pull Request

Description

This PR introduces a v3 LightGBM-based solar power prediction model with enhanced feature engineering, addressing the New Model Challenge (Issue #30).

Why LightGBM?

• Faster training than XGBoost for large datasets
• Better categorical handling with native support
• Leaf-wise tree growth for improved accuracy
• Lower memory usage during training

New Files

Feature Engineering Improvements

1. Cyclical Time Encoding

Handles time discontinuities properly (no more 23→0 hour jumps):

hour_sin = np.sin(2 * np.pi * hour / 24)
hour_cos = np.cos(2 * np.pi * hour / 24)
day_sin = np.sin(2 * np.pi * day_of_year / 365)
day_cos = np.cos(2 * np.pi * day_of_year / 365)

2. Panel Features

• orientation_sin/orientation_cos - Handles 0°/360° wrap-around
• tilt_factor - Effective area calculation based on panel tilt

3. Weather Features

• Temperature, cloud cover (low/mid/high layers)
• Direct normal and diffuse radiation
• Wind speed and precipitation

Training Script Usage

# Install dependency
pip install lightgbm

# Train with sampling (fast development)
python scripts/train_v3_model.py --sample-sites 20 --sample-days 7

# Full training (requires HuggingFace auth for uk_pv dataset)
huggingface-cli login
python scripts/train_v3_model.py --sample-sites 500 --sample-days 365

Initial Results (synthetic data)

• MAE: ~0.23 kW
• Normalized MAE: ~7.7%

Note: Final benchmarks against v1/v2 models will be added once trained on real HuggingFace data.

How Has This Been Tested?

• Training script runs successfully
• Model saves and loads correctly
• Predictions output correct format

Addresses #30

Checklist:

• My code follows OCF's coding style guidelines
• I have performed a self-review of my own code
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• I have checked my code and corrected any misspellings