Online deep learning micrometeorological data generation method based on space-time embedding

Abstract

The invention discloses an on-line deep learning micrometeorological data generation method based on space-time embedding, and the method employs an on-line updating and incremental training mode to fuse a WRF model, a PALM model and a deep learning model into a frame, enables a space-time downscale deep learning model to participate in the operation process of WRF and PALM control programs, and achieves the real-time processing of the WRF and PALM control programs. And learning the change characteristics of the required meteorological elements in the downscaling process, and realizing linkage operation and data cooperative processing of the three models. 3DEST is introduced, so that the model can sense the influence of different geographic positions and height differences on meteorological elements, and the space generalization ability of the model under the conditions of different areas and different heights is remarkably improved. By establishing a unified signal network and data transmission mechanism, continuous generation and automatic acceleration of high-resolution micrometeorological data are realized. According to the method, the problem that in an existing downscaling method based on WRF and PALM, data generation and model training are separated, and the model input size is not uniform, so that the downscaling mode is not universal is solved.

Images