Integrating graph neural networks and attention mechanisms for high-accuracy precipitation prediction

Zhang, Ting (2025) Integrating graph neural networks and attention mechanisms for high-accuracy precipitation prediction. PhD thesis, UTAR.

PDF Download (5Mb)

Abstract

With the advent of big data and meteorological modernization, the role of meteorological data analysis in various fields, such as daily life and social production, has become increasingly prominent. Taking the Guangxi Meteorological Information Center as an example, the volume of meteorological observation data is enormous, posing challenges in data storage, processing, and real-time throughput. Traditional methods of meteorological data analysis are no longer sufficient to deal with the complexity and scale of big data, necessitating the adoption of new approaches to enhance the efficiency and accuracy of weather forecasting. This thesis addresses these challenges by integrating innovative methodologies to organize and analyze meteorological data in Guangxi, focusing on big data processing, neural networks, and other advanced techniques. The research makes three significant contributions: 1. Development of GraphAT-NET: A novel forecasting model named iv GraphAT-NET is proposed, which combines graph neural networks and channel attention mechanisms to enhance short-term precipitation forecasting. By extracting key information from TRAJGRU's feature maps and constructing a Graph structure with an Efficient Channel Attention (ECA) mechanism, GraphAT-NET achieves superior performance. Experiments using the moving MNIST dataset and real-world radar echo data show that GraphAT-NET reduces Mean Squared Error (MSE) to 1.23 and improves the Structural Similarity Index (SSIM) by an average of 16.26% compared to other models. This model demonstrates a substantial average enhancement of 65.4% in MSE and a 10.29% improvement in SSIM on real-world datasets, highlighting its effectiveness in predicting cumulonimbus cloud distribution. 2. Application of the Informer Model: The research explores the use of the Informer model for time series analysis of ground station data. The Informer model achieves a mean absolute error (MAE) of 0.0077, significantly outperforming traditional numerical weather prediction (NWP) techniques with an MAE of 0.02. This deep learning-based framework combines feature engineering and model optimization, enhancing the precision and reliability of short-term precipitation forecasting. The Informer model improves predictive accuracy by 74.2% over conventional methods and by 72.3% over other deep learning-based models, as measured by the root mean square error (RMSE). v 3. Integration and Comparison of Methods: The study integrates GraphAT-NET and the Informer model to forecast rainfall using CAPPI data and ground station data, respectively. By comparing these methods, the results indicate that predictions based on ground station data using the Informer model are more accurate, outperforming radar reflectivity signal-based predictions by 5.16. This highlights the effectiveness of ground station data in precipitation forecasting. In conclusion, this research significantly advances the field of meteorology by introducing innovative deep learning models that improve the accuracy and reliability of short-term precipitation forecasting. The findings have the potential to enhance meteorological services, aiding in better decision-making for weather-related activities and emergency planning. In the future, we will focus on integrating multimodal data for rainfall prediction, combining radar, satellite imagery, ground sensors, and meteorological models. This comprehensive approach aims to enhance prediction accuracy and robustness. We also plan to develop a unified evaluation framework to assess multimodal models fairly, considering each modality's unique characteristics. This will help identify effective approaches, advancing rainfall prediction and improving meteorological services for better decision-making.

Actions (login required)