A Combined Model to Predict GNSS Precipitable Water Vapor Based on Deep Learning

The precipitable water vapor (PWV) is a key parameter to reflect atmospheric water vapor, which can be derived by the global navigation satellite system (GNSS) technique with high accuracy and temporal resolution. PWV is an important parameter for weather forecasts and climate research. To develop a highly accurate PWV prediction model, we first combine the wavelet analysis (Wa), long short-term memory (LSTM) neural network, and autoregressive integrated moving average (ARIMA) algorithms as WLA model for the GNSS PWV prediction. Wa, LSTM, and ARIMA in WLA separate the random noise and predict the nonlinear and linear trends in PWV, respectively. Afterward, the WLA model is compared with LSTM, ARIMA, wavelet neural networks, and the multivariable linear regression (MLR) method. The WLA model shows the best result in the five prediction models in terms of the root-mean-square error (RMSE, 0.19-0.82 mm) and mean absolute error (0.01-0.07), which are 55.48% and 55.32% lower than other models, and Nash-Sutcliffe efficiency coefficient (NSE, 76.53%-99.7%) is 9.42% greater than other models. For further analysis, we also study the WLA performance in different months using one-month's data as training length. The result shows that WLA has good effects in predicting PWV in different months and the average NSE of WLA is 95%. In addition, the predicted PWV of the WLA model within 3 h is found to be accurate and reliable (RMSE < 2 mm, relative error < 0.1, NSE > 60%). This study demonstrates the good performance of WLA to predict GNSS PWV.