A Novel Modeling Strategy of Weighted Mean Temperature in China Using RNN and LSTM

In the meteorology of Global Navigation Satellite System, the weighted mean temperature (Tm) is a key parameter in the process of converting the zenith wetness delay into precipitable water vapor, and it plays an important role in water vapor monitoring. In this research, two deep learning algorithms, namely, recurrent neural network (RNN) and long short-term memory neural network (LSTM), were used to build a high-precision weighted mean temperature model for China using their excellent time series memory capability. The model needs site location information and measured surface temperature to predict the weighted mean temperature. We used data from 118 stations in and around China provided by the Integrated Global Radiosonde Archive from 2010 to 2015 to train the model and data from 2016 for model testing. The root mean square error (RMSE) of the RNN_Tm and LSTM_Tm models were 3.01 K and 2.89 K, respectively. Compared with the values calculated by the empirical GPT3 model, the accuracy was improved by 31.1% (RNN_Tm) and 33.9% (LSTM_Tm). In addition, we selected another 10 evenly distributed stations in China and used the constructed model to test the prediction capability of the weighted mean temperature from 2010 to 2016. The RMSE values were 2.95 K and 2.86 K, which proved that the model also exhibits high generalization in non-modeling sites in China. In general, the RNN_Tm and LSTM_Tm models have a good performance in weighted mean temperature prediction.

Automated summary

In this study, a high-precision weighted mean temperature model for China was built using deep learning algorithms, which showed good accuracy and generalization in prediction. The recurrent neural network (RNN) and long short-term memory neural network (LSTM) models demonstrated strong performance in predicting weighted mean temperature.