ARIMA-KOA-CNN-GRU-Attention Model for Improving GNSS Water Vapor Prediction
TL;DR
Experimental results demonstrated that decomposing PWV into linear and nonlinear components via wavelet transform can enhance prediction accuracy and among all the schemes tested, ARIMA-KOA-CNN-GRU-Attention performed the best, achieving a 52% RMSE reduction in comparison with the ARIMA-KOA-GRU method.
Abstract
Precipitable water vapor (PWV), retrievable from Global Navigation Satellite Systems (GNSS) measurements, is a key indicator of tropospheric water vapor content and crucial for weather forecasting and climate research. However, the prediction of PWV usually relies on a single model or a simple hybrid model, with limited attention to integrated linear and nonlinear characteristics. Therefore, this study proposes a novel hybrid model through integrating both linear and nonlinear components of PWV in the prediction process. Hourly GNSS-PWV and feature parameters including zenith total delay, zenith wet delay, temperature, and atmospheric weighted mean temperature (Tm) are used as its inputs. In the new model, a number of advanced algorithms such as wavelet transform, Kepler optimization algorithm (KOA), aut oregressive integrated moving average (ARIMA), convolutional neural networks (CNN), gated recurrent units (GRU), and attention mechanism were incorporated for the study. Experimental results demonstrated that decomposing PWV into linear (by ARIMA predicted) and nonlinear (by GRU predicted) components via wavelet transform can enhance prediction accuracy. The ARIMA-KOA-GRU model reduced the root mean square error (RMSE) by 19% compared with the KOA-GRU model. Among all the schemes tested, ARIMA-KOA-CNN-GRU-Attention performed the best, achieving a 52% RMSE reduction in comparison with the ARIMA-KOA-GRU method. In addition, those models employing GRU structures outperformed those using long short-term memory neural network structures from all test schemes.