Forecasting Models

Welcome to the main models section of the fusionlab-learn user guide. This section provides detailed documentation for the advanced time series forecasting architectures available in the library.

The models are categorized based on their core architectural principles, from purely data-driven engines to sophisticated hybrids that integrate physical laws.

• Hybrid Models combine the strengths of Recurrent Neural Networks (LSTMs) for sequential processing with the power of Transformers for capturing long-range dependencies.

• Physics-Informed Neural Networks (PINNs) represent a cutting-edge approach, fusing data-driven models with the governing equations of physical systems to produce more robust and consistent results.

• Transformer-Based Models leverage the attention mechanism as their core component, including both pure transformer architectures and variants of the influential Temporal Fusion Transformer (TFT).

The BaseAttentive model serves as the powerful, modular foundation for many of the advanced hybrid and PINN architectures.

Please select a category below to explore the available models.

Model Categories:

• Base Attentive Model Architecture
  • Key Features
  • 1. Initial Feature Processing
  • 2. The Encoder Path
  • 3. The Decoder Path & Context Preparation
  • 4. The Attention Stack
  • 5. Final Output Generation
• Hybrid Models
  • HALNet (Hybrid Attentive LSTM Network)
  • Hybrid Transformer Models: XTFT & SuperXTFT
• Physics-Informed Neural Networks (PINNs)
  • Hybrid Physics-Data Models: PiHALNet & PIHALNet
  • TransFlowSubsNet: A Physics-Informed Hybrid Forecasting Model
  • Physics-Informed Transient Groundwater Flow (PiTGWFlow)
• Transformer-Based Models
  • Pure Transformer Models
  • Temporal Fusion Transformer (TFT) and Variants