A zenith wet delay improved model in China based on GPT3 and random forest

Abstract

Zenith wet delay (ZWD) is a key parameter for the precise positioning of global navigation satellite systems (GNSS) and occupies a central role in meteorological research. Currently, most models only consider the periodic variability of the ZWD, neglecting the effect of nonlinear factors on the ZWD estimation. This oversight results in a limited capability to reflect the rapid fluctuations of the ZWD. To more accurately capture and predict complicated variations in ZWD, this paper developed the CRZWD model by a combination of the GPT3 model and random forests (RF) algorithm using 5-year atmospheric profiles from 70 radiosonde (RS) stations across China. Taking the external 25 test stations data as reference, the root mean square (RMS) of the CRZWD model is 29.95 mm. Compared with the GPT3 model and another model using backpropagation neural network (BPNN), the accuracy has improved by 24.7% and 15.9%, respectively. Notably, over 56% of the test stations exhibit an improvement of more than 20% in contrast to GPT3-ZWD. Further temporal and spatial characteristic analyses also demonstrate the significant accuracy and stability advantages of the CRZWD model, indicating the potential prospects for GNSS-based applications.

Keywords

• Zenith wet delay
• CRZWD model
• GPT3
• Random forest
• Back propagation neural network