⚡ electrical engineering

A Multi-Step Typhoon Track Prediction Method Integrating Spatiotemporal Trajectory Dynamics and Prior Structural Relations

This paper proposes TRGF-KG, a multi-step typhoon track prediction method that integrates spatiotemporal trajectory dynamics with prior structural relations via a dual-branch Transformer and R-GCN architecture to mitigate error accumulation and improve long-horizon prediction accuracy.

Original authors: Chaoran Zhou, Zhuoyang Cui, Hang Chen, Xin Zhang

Published 2026-09-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chaoran Zhou, Zhuoyang Cui, Hang Chen, Xin Zhang

Original paper licensed under CC BY 4.0 (https://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

Typhoons are massive, swirling storms that form over warm ocean waters, capable of unleashing devastating winds, torrential rain, and flooding that threaten lives and property across coastal regions. Because these storms move in complex, shifting patterns driven by the atmosphere and the sea, predicting exactly where they will go days in advance is one of the most difficult challenges in meteorology. While computers have become very good at spotting these storms, forecasting their path more than a few hours into the future often leads to errors that grow larger with time. A small mistake in the beginning can send a predicted path hundreds of miles off course by the end of a week, leaving emergency planners with unreliable information. To solve this, scientists are looking for ways to combine the raw data of a storm's movement with a deeper understanding of the storm's history and structure, hoping to create a more stable and accurate forecast that can hold up over longer periods.

Researchers at Changchun University of Science and Technology and the Institute of Scientific and Technical Information of Jilin have developed a new method called TRGF-KG to tackle this problem of long-range typhoon prediction. Instead of relying on a single type of data, their approach acts like a dual-engine system that processes two different kinds of information simultaneously. The first engine looks at the storm's recent history as a continuous stream of numbers, tracking its speed, pressure, and location step-by-step to understand how it is moving right now. The second engine, however, treats the storm's history not just as a list of numbers, but as a structured map of relationships. This map connects the storm to its specific type, the data sources that recorded it, and the sequence of observations that show how it evolved over time. By building this structured map, known as a knowledge graph, the researchers give the computer a way to "remember" the logical connections between different parts of the storm's life, rather than just seeing a blur of past positions.

The core innovation lies in how these two engines work together. In many previous attempts, scientists simply tried to mix these two types of data by stacking them together, which often caused confusion because the data spoke different "languages." The new method uses a smart filtering system to align the structured map with the moving stream of data. It checks the time stamps of the storm's history to ensure the map and the movement data match up perfectly at every step. Then, it uses a gating mechanism—a kind of adjustable valve—to decide how much influence the structured map should have on the final prediction at any given moment. This allows the system to lean heavily on the storm's immediate movement when the path is clear, but to pull in the structural knowledge when the storm behaves unpredictably, effectively correcting the forecast before it drifts too far off course.

To test this idea, the team trained their system on a massive dataset of typhoon tracks from the western Pacific Ocean, covering records from the year 2000 to 2022. They compared their new method against several other advanced computer models, including those that only look at movement data and those that use different ways of organizing knowledge. The results showed that the new method consistently outperformed the others. When predicting the storm's location 24 steps into the future, which corresponds to a forecast 144 hours ahead, the new system reduced the average distance error to about 275 kilometers. This was significantly better than the next best model, which had a larger error. The improvement was not just a small tweak; the new method maintained a much higher level of accuracy as the forecast horizon extended, proving that adding the structured knowledge helped prevent the errors from piling up over time.

The researchers also looked at specific real-world examples to see how the method handled different types of storms. They tested it on Typhoon Chanthu, which made a sharp, difficult turn, and Typhoon Muifa, which moved in a smoother, more predictable line. In the case of the sharp turn, older models tended to drift away from the actual path, failing to capture the sudden change in direction. The new method, however, stayed much closer to the true track, successfully navigating the curve. Even for the smoother storm, the new system provided a more precise forecast than the alternatives. These findings suggest that by explicitly teaching the computer about the relationships and structure behind a storm's history, rather than just its raw movement, scientists can create forecasts that remain reliable for longer periods, offering a crucial advantage for disaster preparedness and safety.

Technical Summary: TRGF-KG for Multi-Step Typhoon Track Prediction

Problem Statement

Accurate multi-step prediction of typhoon tracks is critical for disaster preparedness but remains challenging due to the nonlinear, time-varying, and uncertain nature of typhoon motion. Existing deep learning approaches, such as LSTM and Transformer-based models, primarily treat typhoon evolution as continuous sequences. While effective at capturing temporal dependencies, these methods often suffer from error accumulation over long prediction horizons and fail to explicitly model the underlying structural relations among typhoon events, observations, and temporal connections. Furthermore, simply concatenating graph-derived representations with trajectory features often leads to semantic and temporal mismatches, limiting the effectiveness of integrating prior structural knowledge.

Methodology: TRGF-KG

The paper proposes TRGF-KG (Trajectory–Relation Graph Fusion with Knowledge Graph), a dual-branch framework designed to integrate spatiotemporal trajectory dynamics with prior structural relations derived from a typhoon knowledge graph.

1. Overall Architecture

The framework consists of three main components:

Block-Level Dynamic Temporal Representation Branch: Uses a Transformer encoder to process historical observation sequences (longitude, latitude, wind speed, pressure). The input sequence is divided into non-overlapping temporal blocks to capture both local evolutionary variations and cross-block long-range dependencies.
Relational Structural Prior Representation Branch: Constructs a typhoon knowledge graph (KG) based on an ontology linking disaster categories, typhoon events, data sources, and time-indexed observation nodes. A Relational Graph Convolutional Network (R-GCN) with basis decomposition learns relational structural prior representations from this graph, explicitly encoding event affiliation, observation inclusion, and temporal evolution.
Fusion and Prediction Module: A novel mechanism that aligns and fuses the outputs of the two branches.

2. Key Technical Mechanisms

Time-Index-Based Cross-Modal Collaborative Alignment: To address the mismatch between graph-derived representations and trajectory blocks, the KG representations are transformed and aligned with the temporal blocks using typhoon identifiers and time indices. This ensures that relational priors correspond correctly to specific time steps in the trajectory.
Feature-Wise Gated Residual Fusion: Instead of simple concatenation, the model employs a gating mechanism to selectively inject relational structural priors into the temporal backbone.
- Global average pooling is applied to both modalities, and a gating network generates hidden-channel-level weights ( $g$ ).
- The fused representation is calculated as: $H_{fused} = H_{trad} + g \odot \text{MLP}_{trans}(H'_{kg})$ .
- This design preserves the temporal branch as the prediction backbone while adaptively incorporating relational information to mitigate noise and error accumulation.

3. Training and Evaluation

The model is trained as a supervised regression task using Mean Squared Error (MSE) on standardized data. Evaluation is performed on the Chinese Coastal Typhoon Track Dataset (2000–2022) using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), Coefficient of Determination ( $R^2$ ), and Average Distance Error (ADE).

Key Contributions

Dual-Branch Framework: The proposal of TRGF-KG, which jointly models temporal trajectory dynamics and knowledge-graph-derived prior structural relations, addressing the limitation of sequence-only models.
Novel Fusion Strategy: The development of a time-index-based cross-modal alignment strategy and a feature-wise gated residual fusion mechanism. This allows the model to adaptively incorporate relational representations while maintaining the continuity and dominant role of trajectory dynamics.
Empirical Validation: Comprehensive experiments demonstrating that integrating explicit structural priors improves prediction accuracy and long-horizon stability compared to temporal-only and other knowledge-graph-enhanced baselines.

Experimental Results

Experiments on the Chinese Coastal Typhoon Track Dataset show that TRGF-KG outperforms various baselines, including LSTM, Transformer, TransE-KG, TransR-KG, GAT-KG, and other KG-enhanced temporal backbones (TimesNet-KG, SegRNN-KG, ETSformer-KG).

Accuracy: TRGF-KG achieved the lowest MAE and RMSE for both longitude and latitude predictions. For longitude, it achieved an MAE of 1.0261° and $R^2$ of 0.9384; for latitude, an MAE of 0.9014° and $R^2$ of 0.9505.
Long-Horizon Stability: The model maintained the highest $R^2$ values at the 24-step horizon (0.8587 for longitude, 0.8789 for latitude), indicating superior stability in reducing error accumulation over extended prediction periods.
Spatial Deviation: TRGF-KG achieved the lowest Average Distance Error (ADE) of 161.8773 km, significantly outperforming the next best model (TimesNet-KG).
Ablation Studies: Results confirmed that the performance gain is not merely due to adding graph data but specifically due to the feature-wise gated residual fusion. A variant with shared gating (Baseline+RGCN+SG) performed worse than the full model, highlighting the necessity of feature-level adaptive fusion.
Case Studies: Visualizations of Typhoon "Chanthu" (complex turning) and "Muifa" (smooth track) demonstrated that TRGF-KG better reconstructs curved evolution and suppresses spatial bias compared to other models.

Significance and Claims

The paper claims that TRGF-KG effectively addresses the vulnerability of multi-step typhoon track prediction to error accumulation and track deviation. By treating the typhoon knowledge graph as a relational structural prior, the method explicitly preserves event affiliation and temporal ordering that are difficult to represent from continuous sequences alone.

The authors assert that the primary significance lies in the aligned and adaptive integration of these structural priors. This approach allows relational information to complement temporal dynamics without introducing noisy interactions, thereby improving the accuracy and long-horizon stability of predictions. The study concludes that explicit relational structural representations provide an effective complement to continuous temporal modeling, offering a robust solution for multi-step typhoon track prediction. Future work is noted to focus on incorporating richer meteorological and physical information into the knowledge graph.