maml-stgcn-lstm-weather

Hybrid MAML-STGCN-LSTM for Global-to-Regional Weather Forecasting (Model v5.0)

Overview

This repository contains Model v5.0, a meta-learning–based spatio-temporal graph neural network for weather forecasting, designed to generalize across global climate regions and rapidly adapt to new, unseen locations using limited local data.

The model integrates:

• Spatio-Temporal Graph Convolutional Networks (STGCN) for spatial dependency modeling
• Deep multi-layer LSTM for long-range temporal forecasting
• Model-Agnostic Meta-Learning (MAML) to enable fast regional adaptation

The system is trained on globally distributed climate regions and fine-tuned to specific regions (e.g., deserts, tropics, polar zones) using multi-year historical data.

---

Model Architecture

Core Components

• STGCN Backbone
  • Captures spatial dependencies between geographic grid points
  • Graph constructed using k-nearest-neighbor connectivity
• LSTM
  • Deep LSTM stack for temporal sequence modeling
  • Improves long-horizon forecast stability
• MAML Meta-Learning Framework
  • Learns initialization parameters that adapt quickly to new regions
• Köppen Climate Embedding
  • Climate-type–aware embeddings injected into model inputs

Model Scale

---

Dataset

ERA5 Reanalysis (Single Levels)

This model is trained using the ERA5 hourly single-levels reanalysis dataset from cds.climate.copernicus.eu

Dataset characteristics

• Temporal resolution: Hourly
• Spatial resolution: 0.25° × 0.25° (30 km)
• Coverage: Global
• Data type: Surface and single-level atmospheric variables
• Format: NetCDF

ERA5 provides physically consistent and globally complete weather fields, making it suitable for spatio-temporal learning and climate generalization tasks.

---

Training Data Details

Meta-Training (Global Training)

• Number of regions: 15 globally distributed regions
• Climate coverage:
  • Tropical, temperate, desert, polar, and monsoon climates
• Sampling strategy:
  • Each region is treated as an independent meta-learning task
• Per-region data usage:
  • Up to ~600 spatio-temporal samples per region
  • Support / query split:
    • ~75% support (inner-loop adaptation)
    • ~25% query (meta-update)

Each training sample consists of:

• A 24-step spatio-temporal input window
• An 8-step forecast horizon
• A graph-structured spatial representation

---

Regional Adaptation Data

After meta-training, the model is adapted to new regions using multi-year historical data.

• Years used: 2023–2024
• Seasonal coverage: All four quarters per year
• Sampling cap: Up to ~1,200 windowed samples per region
• Train / validation split:
  • 80% training
  • 20% validation

Each training sample consists of:

• A 24-step spatio-temporal input window
• An 8-step forecast horizon

---

Regional Adaptation Data

After meta-training, the model is adapted to new regions using multi-year historical data.

• Years used: 2023–2024
• Seasonal coverage: All four quarters per year
• Sampling cap: Up to ~1,200 windowed samples per region
• Train / validation split:
  • 80% training
  • 20% validation

Training Strategy

Meta-Learning (MAML)

• Inner Loop
  • SGD-based fast adaptation
  • Multiple inner epochs per task
• Outer Loop
  • AdamW optimizer
  • Gradient accumulation
  • Cosine annealing with warm restarts
• Stability techniques
  • Gradient clipping
  • Adaptive task sampling based on task difficulty

---

Regional Adaptation

The pretrained model has been adapted and evaluated on diverse regions, including:

MODEL_REGIONS = [
    (18, 23, 75, 80),  # India
    (8, 13, 98, 103),  # Thailand
    (53, 58, 35, 40),  # Russia
    (12.5, 17.5, 102.5, 107.5),  # Thailand/Cambodia
    (22.5, 27.5, 19.5, 24.5),  # Libya/Egypt
    (43.5, 48.5, 7.5, 12.5),  # Southern France
    (35.5, 40.5, -5.5, -0.5),  # Spain/Mediterranean
    (32.5, 37.5, 137.5, 142.5),  # Tokyo/Eastern Japan
    (-23.5, -18.5, 132.5, 137.5),  # Australia
    (-20, -15, -70, -65),  # Peru
    (44.5, 49.5, 125.5, 130.5),  # Northeast China
    (29.5, 34.5, -101.5, -96.5),  # Texas
    (-9.5, -4.5, -67.5, -62.5),  # Amazon Basin
    (67.5, 72.5, -32.5, -27.5),  # Greenland
    (51.5, 56.5, -112.5, -107.5),  # Alberta, Canada
]

Adaptation characteristics:

• Multi-year seasonal data
• Climate-aware optimizer and learning-rate scheduler
• Full fine-tuning of STGCN and LSTM layers
• Optimized for fast convergence

---

Evaluation

• Loss function: Mean Squared Error (MSE)

• Validation performed on held-out regional datasets

---

Limitations

• Computationally intensive due to GNN + LSTM + MAML architecture

• Requires gridded weather data in NetCDF format

• Uses static spatial graphs

• Not optimized for real-time inference

Research Motivation

This project was developed as an academic research-oriented system focusing on:

• Meta-learning for spatio-temporal forecasting

• Climate-aware adaptation mechanisms

• Graph-based representations of physical systems

The design emphasizes extensibility and reproducibility, making it suitable for research papers, final-year projects, and graduate-level coursework.

Author

Yashas Gowda N AIML Undergraduate Research interests: Meta-Learning, Graph Neural Networks, Climate AI

License

MIT License