Downscaled Correction of Temperature Forecast Algorithm with Encoder-Decoder over the Hengduan Mountains

Global climate models (GCMs), a primary tool of Numerical weather prediction (NWP) has an irreplaceable position in weather forecasting. Because of the inevitable model error and the coarse resolution, it is necessary to develop the downscaled correction method to reduce the model prediction error and improve the forecast result on the basis of the existing model, especially in the mountains. The Hengduan Mountains (HMs) belongs to the monsoon region accompanied by the highly complex mountainous topography, leading to the accuracy of GCMs limited in this region. The study aims to propose a more accurate deep learning algorithm than the downscaled correction model based on the deep learning method before. In this study, downscaled correction of temperature forecast algorithm with Encoder-Decoder (DCTA-ED) is proposed for downscaling the GCMs results of 2m temperature from coarse grid data to meteorological station point forecast over HMs. The proposed algorithm extracts the interaction between multiple meteorological data and time features of temperature, wind speed, air pressure and humidity from previous position observation, to make the downscaled correction prediction results more consistent with the characteristics of positions. The experimental results demonstrate that the proposed algorithm exceeds most of the downscaled correction models for weather stations with different terrains.