Refined Short‐Term Forecasting Atmospheric Temperature Profiles in the Stratosphere Based on Operators Learning of Neural Networks

Abstract

The efficacious forecasting of single‐station atmospheric temperature profiles can provide essential support for the structural design and flight missions of spacecrafts in near space. However, empirical models and reference atmospheric models most are calculations of the average state of the atmosphere profiles. Numerical assimilation models require expensive computational costs to improve the accuracy for medium and long‐term forecasting. It has been still a challenge to refined predict short‐term temperature profiles of near space at a low‐cost. We present a temperature profile operator method for refined modeling in the stratosphere by fusing the ability of Long Short‐Term Memory (LSTM) networks or its variants‐ bidirectional LSTM (BiLSTM) to exploit time series correlated information and deep operator networks (DeepONets) to approximate the solution operator of temperature profiles. It consists of three subnetworks. The first subnetwork is used to approximate the discrete temperature profile function, the second net is applied to represent the spatial information of pressure heights, and the third branch is utilized to encode the time domain of the temperature profile operator. We first use the hourly low latitude temperature data (20–50 km) from ERA5 for training, verification and iterative testing in the next 48 hr. The results denote that the temperature profile operator network has a stable and low error of cumulative generalization, and the BiLSTM operator significantly outperform the other models. We also apply two scenarios to testing the refined applicability of the high latitude temperature profile operator and the mid latitude wind profile operator in the stratosphere. This work provides a novel perspective for us to study the refined single‐station modeling of the upper and middle atmosphere.