ISSN 1000-0526
CN 11-2282/P
CN 11-2282/P
Archive > Volume 52 Issue 5 > 2026,52(5):538-551. DOI:10.7519/j.issn.1000-0526.2026.030903 Prev Next
Temperature Forecasting Method of Physics-Constrained Deep Learning Integrated with Pangu-Weather Model Forecast Products
- Article
- Figures
- Metrics
- Preview PDF
- Reference
- Related
- Cited by
- Materials
Abstract:
Aiming at the fine-scale forecasting challenge of 2 m temperature (T2m) in complex terrain areas, this paper selects the Guangxi Region, a typical area with complex terrain, as the research object and proposes a physics-constrained deep learning forecasting model named PSD-Net, which integrates the forecast products of the Pangu-Weather Model. The forecast products of Pangu-Weather (PANGU) are used as feature variables input. The generator based on the super-resolution generative adversarial network is employed to extract multi-scale features. Power-spectral-density and Kullback-Leibler divergence are explicitly injected into the loss function as constraint terms so as to improve the consistency of forecast products with observations in spectral fidelity and probability distribution. Compared to the T2m forecast performance of the ECMWF, SCMOC and PANGU products in the Guangxi Region in 2024, both the grided forecasts and station-based forecasts of PSD-Net outperform the compared forecast products. In particular, the mean absolute error (MAE) of the gridded forecasts is reduced by 37.6% relative to that of PANGU and the accuracy is improved by 17 percentage points. The growths of both MAE and root mean square error (RMSE) for the 1-72 h T2m forecast products from PSD-Net are less than those of the compared forecast products, and there is a gentle error growth at lead time 25-72 h. In a word, this study has verified the effectiveness of the physics-constrained deep learning framework in fine-scale T2m forecasting which could provide a new approach for the combination of meteorological and AI models.
Keywords:
Project Supported:
Clc Number:
Mobile website
