Research on GNSS elevation time series prediction based on STL-XGBoost model

In order to enhance the accuracy of GNSS elevation time series prediction and address the limitations of conventional GNSS elevation time series prediction methods, which include imperfect feature selection and poor stability, this paper proposes a novel prediction approach based on a STL-XGBoost combined model. This method employs STL decomposition to segregate GNSS data into trend and non-trend components. The non-trend items are then utilized as the input data for the time-characterized XGBoost model, which employs its robust nonlinear modeling capabilities to predict GNSS elevation time series. Concurrently, STL decomposition is employed to forecast the trend item. The nonlinear predicted value and the trend item predicted value are then combined to yield the final combined model predicted value. The analysis and verification of multipl e GNSS data sets revealed a 32% reduction in the mean MAE and a 30% reduction in the mean RMSE for the STL -XGBoost model prediction results, in comparison with the single STL and XGBoost model prediction results. These findings indicate enhanced accuracy and a robust correlation with the original time series. The combined model proposed in this paper has markedly enhanced the precision and reliability of predictions, offering an efficacious approach to analysing GNSS elevation time series.