🤖 machine learning

Leveraging Teleconnections with Physics-Informed Graph Attention Networks for Long-Range Extreme Rainfall Forecasting in Thailand

This paper proposes a novel physics-informed Graph Attention Network combined with a spatial season-aware Generalized Pareto Distribution to improve long-range extreme rainfall forecasting in Thailand by leveraging teleconnections and spatiotemporal graph structures.

Original authors: Kiattikun Chobtham, Kanoksri Sarinnapakorn, Kritanai Torsri, Prattana Deeprasertkul, Jirawan Kamma

Published 2026-04-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kiattikun Chobtham, Kanoksri Sarinnapakorn, Kritanai Torsri, Prattana Deeprasertkul, Jirawan Kamma

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Weather Whisperer: How Math and Physics Help Predict Thailand's Big Storms

Imagine you are trying to predict if a massive wave will hit a specific beach in Thailand months from now. You can’t just look at the sand on that beach; you have to look at the wind in the middle of the ocean, the temperature of the water thousands of miles away, and even the shape of the mountains nearby that might push the water upward.

Predicting "extreme rainfall"—those massive, scary downpours that cause floods—is one of the hardest jobs in science. Most computer models are either too "general" (like looking at a map from a satellite) or too "blind" (they see patterns but don't understand why they happen).

This paper introduces a new way to predict rainfall in Thailand using a "smart" system that combines Physics, Social Networks, and Extreme Math.

1. The "Social Network" of Rain (Graph Attention Networks)

Think of every rain gauge station in Thailand as a person in a giant social network.

The Connections: In a normal social network, you are connected to your friends. In this model, rain stations are "connected" to each other based on teleconnections. For example, if a change in ocean temperature in the Pacific Ocean (like El Niño) always causes rain in Thailand a month later, those two "people" (the ocean and the station) are "friends" in our network.
The Attention Mechanism: Not all friends are equally important. If you want to know if it will rain tomorrow, your "best friend" (a nearby station or a major ocean pattern) is more important than a "distant acquaintance." The model uses something called "Attention" to decide which "friends" to listen to most closely when making a prediction.

2. Giving the AI a "Physics Textbook" (Physics-Informed)

Most AI models are "black boxes"—they look at data and guess, but they don't actually understand the world. They might see a pattern that is actually impossible in real life.

The researchers fixed this by giving the AI a "physics textbook." They taught the model about Orographic Precipitation.

The Analogy: Imagine wind blowing air toward a mountain. As the air climbs the mountain, it gets squeezed and cooled, turning into rain.
Instead of letting the AI guess, the researchers gave it the mathematical rules of how mountains and wind interact. This ensures the AI's "guesses" are grounded in the real laws of nature.

3. The "Safety Net" for Extremes (Spatial Season-aware GPD)

Standard AI models are like students who are good at passing average tests but fail when things get "extreme." Because massive floods are rare, the AI doesn't see them often enough to learn them perfectly. It tends to "play it safe" and predict average rain, which is dangerous because it misses the life-threatening floods.

To fix this, the researchers added a "Safety Net" called GPD mapping.

The Analogy: Imagine a professional athlete. They can run a steady pace for miles, but when it’s time for a 100-meter sprint, they need a different kind of energy.
The GPD is a special mathematical tool that looks specifically at the "tails" of the data—the rare, extreme moments. It takes the AI's "average" prediction and says, "Wait, based on the season and the location, this has the potential to be a massive storm!" It adjusts the prediction upward to make sure the "big ones" aren't ignored.

Why does this matter?

In the real world, this isn't just about math; it's about survival and planning.

By comparing their system to the world-standard models (like the ones used by the European weather centers), the researchers found that their method was much better at predicting the actual amount of rain in Thailand.

The Result: Better maps and better warnings. This helps farmers know when to plant, helps city planners prepare drainage systems, and helps dam managers decide how much water to release before a storm hits. It turns "guessing the weather" into "understanding the Earth."

Technical Summary: Leveraging Teleconnections with Physics-Informed Graph Attention Networks for Long-Range Extreme Rainfall Forecasting in Thailand

1. Problem Statement

Accurate long-range rainfall forecasting (1–12 months ahead) is critical for water resource management and agricultural planning. However, traditional methods face two primary challenges:

Numerical Weather Prediction (NWP) Limitations: Physics-based models like ECMWF’s SEAS5 are computationally expensive, require significant expertise, and often produce coarse-resolution outputs that are difficult to apply to local decision-making.
Machine Learning (ML) Limitations: While deep learning models are computationally efficient, they often act as "black boxes" lacking interpretability. Furthermore, standard ML models struggle to predict extreme rainfall events because they tend to underestimate the "tails" of probability distributions due to the scarcity of extreme observations.

2. Methodology

The authors propose an end-to-end forecasting framework that integrates atmospheric physics, graph theory, and extreme value statistics. The methodology is divided into three core components:

A. Physics-Informed GAT-detrend (Spatiotemporal Modeling):

Graph Construction: Instead of a regular grid, the model uses a graph structure where gauge stations are nodes. Edges are defined by teleconnections (spatial dependencies) identified through Pearson correlation and Granger Causality tests.
Physics-Informed Edge Features: To prevent the model from ignoring fundamental physics, the authors incorporate the Smith linear model of orographic precipitation. This uses terrain elevation and wind components ( $U$ and $V$ ) to simulate the probability of rainfall, providing "prior knowledge" to the graph's edges.
GAT-detrend Architecture: The model utilizes a Graph Attention Network (GAT) to dynamically weight the influence of neighboring stations and climate indices. To handle non-stationary time-series data, a detrending technique (using Layer Normalization and MLP-based level estimation) is applied to the input features.

B. Spatial Season-aware GPD (Extreme Value Analysis):

To correct the underestimation of extremes, the authors implement a Peak-Over-Threshold (POT) approach using a novel Spatial Season-aware Generalized Pareto Distribution (GPD).
Unlike standard bias correction, this method fits GPD parameters (shape and scale) specifically for each station and its corresponding local season (e.g., SW monsoon, NE monsoon, or dry season). This allows the model to "map" predicted values into the correct extreme-value distribution.

C. Hierarchical Clustering:

The researchers used hierarchical clustering (incorporating PCA and Euclidean distance) to group rainfall stations into 12 distinct spatial clusters, allowing for localized model training that respects regional hydroclimatic regimes.

3. Key Contributions

Hybrid Modeling: Successfully bridges the gap between data-driven GNNs and classical thermodynamic physics (orographic precipitation).
Explainability: By using attention mechanisms on teleconnection-based graphs, the model provides "signal attribution," showing which climate indices (e.g., ENSO, MJO, PDO) drive rainfall in specific regions.
Extreme-Value Correction: Introduces a seasonal-aware GPD mapping technique that specifically targets the tail of the distribution, improving the accuracy of heavy rainfall predictions.
Operational Framework: Provides a complete workflow from raw climate indices to high-resolution interpolated rainfall maps.

4. Results

Superior Accuracy: The GAT-detrend model outperformed established baselines (LSTM, Wavelet-LSTM, GC-LSTM, and EvolveGCN) across most clusters.
Extreme Event Performance: In Southern Thailand (Clusters 1–4), where extreme rainfall is most frequent, the combination of physics-informed features and GPD mapping significantly improved accuracy. For example, in Cluster 4, the model achieved 81.08% accuracy, vastly outperforming EvolveGCN (52.14%).
Error Reduction: The inclusion of GPD mapping and physics markedly lowered the Root Mean Squared Error (RMSE) and improved the Nash–Sutcliffe Efficiency (NSE), indicating a much stronger correspondence with observed rainfall.
Real-world Benchmarking: When compared to the operational ECMWF SEAS5 system, the proposed model achieved a lower average RMSE (by ~21.83 mm/month) and higher overall accuracy (79.3% vs. 72.1%), while providing higher-resolution maps suitable for local water management.

5. Significance

This research demonstrates that incorporating domain-specific physical constraints and extreme value theory into deep learning architectures can overcome the inherent limitations of "pure" data-driven models. The ability to provide interpretable, high-resolution, and extreme-event-aware forecasts makes this framework a highly practical tool for climate-risk management, reservoir operation, and disaster mitigation in monsoon-driven regions like Thailand.