Zenith Troposphere Delay Prediction based on BP Neural Network and Least Squares Support Vector Machine

Abstract

<p><strong>Abstract</strong><strong>:</strong> With the rapid development of artificial intelligence, machine learning has become an high-efficient tool applied in the fields of GNSS data analysis and processing, such as troposphere, ionosphere or satellite clock modeling and prediction. In this paper, zenith troposphere delay (ZTD) prediction algorithms based on BP neural network (BPNN) and least squares support vector machine (LSSVM) are proposed in the time and space domain. The main trend terms in ZTD time series are deducted by polynomial fitting, and the remaining residuals are reconstructed and modeled by BPNN and LSSVM algorithm respectively. The test results show that the performance of LSSVM is better than that of BPNN in term of prediction stability and accuracy by using ZTD products of International GNSS Service (IGS) of 20 stations in time domain. In order to further improve LSSVM prediction accuracy, a new strategy of training samples selection based on correlation analysis is proposed. The results show that using the proposed strategy, about 80% to 90% of the 1-hour prediction deviation of LSSVM can reach millimeter level depending on the season, and the percentage of the prediction deviation value less than 5 mm is about 60% to 70%, which is 5% to 20% higher than that of the classical random selection in different month. The mean values of RMSE in all 20 stations using the new strategy are 1-3mm smaller than those of the classical one. Then different prediction span from 1 to 12 hours is conducted to show the performance of the proposed method. Finally, the ZTD predictions based on BPNN and LSSVM in space domain are also verified and compared using GNSS CORS network data of Hong Kong, China.</p><p><strong>Keywords</strong><strong>:</strong> ZTD, BP Neural Network, Support Vector Machine, Least Squares, GNSS</p><p><strong>Acknowledgments:</strong> This work was supported by Natural Science Foundation of China (41874032) and the National Key Research and Development Program (2016YFB0501701)</p><p> </p>