Research on Short-Term Forecasting Model of Global Atmospheric Temperature and Wind in the near Space Based on Deep Learning

Abstract

Developing short-term forecasting models for global atmospheric temperature and wind in near space is crucial for understanding atmospheric dynamics and supporting human activities in this region. While numerical models have been extensively developed, deep learning techniques have recently shown promise in improving atmospheric forecasting accuracy. In this study, convolutional long short-term memory (ConvLSTM) and convolutional gated recurrent unit (ConvGRU) neural networks were applied to build for short-term global-scale forecasting model of atmospheric temperature and wind in near space based on the MERRA-2 reanalysis dataset from 2010–2022. The model results showed that the ConvGRU model outperforms the ConvLSTM model in the short-term forecast results. The ConvGRU model achieved a root mean square error in the first three hours of approximately 1.8 K for temperature predictions, and errors of 4.2 m/s and 3.8 m/s for eastward and northward wind predictions on all 72 isobaric surfaces. Specifically, at a higher altitude (on the 1.65 Pa isobaric surface, approximately 70 km above sea level), the ConvGRU model achieved a RMSE of about 2.85 K for temperature predictions, and 5.67 m/s and 5.17 m/s for eastward and northward wind. This finding is significantly meaningful for short-term temperature and wind forecasts in near space and for exploring the physical mechanisms related to temperature and wind variations in this region.