Add DLWP (Deep Learning Weather Prediction) for Weather Forecasting

[cristian-victoria]

This PR implements DLWP, a U-Net-based model for global weather forecasting on cubed-sphere grids, as described in Weyn et al. (2020, 2021).

Reference:

Weyn et al. (2021) - Sub-Seasonal Forecasting
Weyn et al. (2020) - Cubed Sphere CNNs

Changes

New Files

• pyhazards/models/dlwp.py - Core implementation
  • DLWP class - U-Net architecture
  • dlwp_builder() - builder function
  • DoubleConv, Down, Up - blocks
  • WeatherMetrics class - RMSE, ACC, Bias metrics
• docs/source/modules/models_dlwp.rst - Full documentation

Modified Files

• pyhazards/models/__init__.py - Added imports and registration
• docs/source/pyhazards_models.rst - Added model to index

Smoke Testing

Results:

Test 1: Model Registration
Test 2: Forward Pass
Test 3: Shape Validation
Test 4: Training Loop
Test 5: Gradient Flow
Test 6: Loss and Metrics
Test 7: Save and Load
Total: 7/7 tests passed

Paper Parity Note

This implementation follows the WaveCastNet paper with these intentional differences:

Identical to Paper

• Architecture:

  • U-Net encoder-decoder with skip connections
  • DoubleConv blocks (Conv → BN → Act × 2)
  • MaxPool downsampling, ConvTranspose upsampling
  • Channel doubling at each level

• Data Format:

  • Cubed-sphere grid: (batch, channels, 6_faces, height, width)
  • Single-step prediction: u(t) → u(t+6hr)
  • Autoregressive multi-step forecasting

• Training & Evaluation:

  • MSE loss function
  • RMSE, ACC, Bias metrics

PyHazards-Specific Adaptations

• Training loop: Uses Trainer.fit() instead of custom loops
• Data format: Uses DataBundle wrapper
• Face processing: Reshapes (B,C,F,H,W) → (B*F,C,H,W) for Conv2d

Future Work

• Mixed precision training (can add via torch.cuda.amp)
• Distributed training (can add via DataParallel)
• Learning rate scheduling (can add via PyTorch schedulers)
• Specialized cubed-sphere convolutions (advanced feature)

These omissions don't affect core functionality and can be added later if needed.