ISSN 1000-0526
CN 11-2282/P
CN 11-2282/P
A Deep Learning-Based Method for Weather Situation Similarity Forecasting: SSN
- Article
- Figures
- Metrics
- Preview PDF
- Reference
- Related
- Cited by
- Materials
Abstract:
Similarity forecasting is a statistical forecasting method widely used in operational services. To overcome the limitations of single-layer similarity, which lacks three-dimensional spatial information and suffers from unstable performance due to the frequent interference of weather system shapes and values, this paper explores the feasibility of deep learning models in weather situation recognition and forecasting. Using ERA5 reanalysis data, we construct a deep learning model based on CNN and Transformer modules, incorporating a self-attention mechanism. The model is trained and optimized, and the test results indicate that it can accurately extract the three-dimensional spatial information of weather situations. We then combine the Pearson correlation, which emphasizes shape, the Euclidean distance, which emphasizes value, and the Chebyshev criterion, which comprehensively considers both shape and value, to design a new weather situation similarity forecasting method, SynopticSimNet (SSN). Finally, we conduct a detailed evaluation of its practical performance. The results show that: (1) the SSN method achieves the highest average SSIM and the lowest MSE, with a significant improvement in SSIM and reduction in MSE compared to traditional methods; (2) Analysis of cases from two different seasons shows that the best historical similar cases identified by SSN are not only numerically closer to the original field but also exhibit the most consistent spatial distribution, demonstrating the method"s promising potential for operational application.
Keywords:
Project Supported:
Clc Number:
Mobile website
