Research on ZWD Forecasting Model Based on Improved Random Forest Algorithm

Abstract When electromagnetic wave signals from the Global Navigation Satellite System (GNSS) pass through the troposphere, they are impeded by atmospheric conditions, thereby affecting positioning efficiency. The atmosphere is rich in water vapor information, which can typically be studied quantitatively through tropospheric zenith wet delay (ZWD). Therefore, high-precision forecasting models of ZWD are of significant research importance in GNSS positioning enhancement, weather forecasting, and water vapor inversion. The Random Forest (RF) algorithm has advantages in terms of improving accuracy, resistance to overfitting, and assessing feature importance. Aiming at the complex spatiotemporal patterns of global ZWD and the difficulty in establishing high-precision models, this paper develops a high-precision ZWD forecasting model based on the RF algorithm. In this study, we processed global ZWD data from 2021 to 2024, studied the spatiotemporal distribution characteristics of ZWD from 2021 to 2023, analyzed the spatial distribution patterns and temporal variation trends of ZWD, and the time series data of ZWD was used as input for the RF algorithm to construct a global ZWD forecasting model, plotted the testing set prediction results, regression charts, error histograms, feature importance diagrams, and error curves, and calculated precision indicators such as mean absolute error, root mean square error, mean percentage error, and coefficient of determination to comprehensively evaluate the model's accuracy. The results indicate that the ZWD forecasting model based on the RF algorithm has a higher accuracy in the equatorial region, with an optimal value reaching 96.14%. When predicting for stations worldwide, the testing set's mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) are all relatively small, while the R2 value is comparatively large, with average values of 1.623 cm, 2.146 cm, 19.478%, and 0.691, respectively. Compared to similar models, the new model developed in this study shows a significantly improved forecasting performance. In the current landscape of ZWD forecasting models, there is a scarcity of models that offer global coverage and extended forecast durations. The model developed in this study is capable of providing one-month ahead forecasts for ZWD across the globe, thereby providing theoretical and data support for the establishment of additional long-term global forecasting models in the future.