Hybrid Models

Welcome to the Hybrid Models section of the fusionlab-learn user guide. The models detailed here represent a powerful architectural paradigm that combines the strengths of different deep learning components to tackle complex time series forecasting tasks.

Specifically, these “hybrid” models fuse the sequential processing capabilities of Recurrent Neural Networks (LSTMs) with the sophisticated context-modeling power of Transformers and their attention mechanisms.

This approach allows the models to:

• Capture short-term temporal dependencies and sequential patterns effectively using LSTMs.

• Model long-range dependencies and complex feature interactions using multi-headed attention.

• Integrate diverse data sources (static, dynamic past, and known future) into a cohesive and rich representation.

This section provides detailed guides for each of the hybrid models available in the library.

Hybrid Models:

• HALNet (Hybrid Attentive LSTM Network)
  • Key Features
  • When to Use HALNet
  • Complete Example
• Hybrid Transformer Models: XTFT & SuperXTFT
  • XTFT (Extreme Temporal Fusion Transformer)
  • SuperXTFT: An Enhanced Hybrid Transformer
  • Next Steps