Accuracy analysis of ionospheric delay regression prediction based on multi-feature BP neural network

An ionospheric delay regression model based on a multi-feature backpropagation neural network (BPNN) is proposed to address the problem of reduced spatial correlation of ionospheric delay in long baseline network real-time dynamic positioning (RTK) technology due to significant differences in ionospheric conditions at both ends of the baseline. After double difference processing, residual ionospheric delay cannot be eliminated, resulting in increased errors and decreased positioning accuracy. The model combines spatiotemporal parameters (local time, baseline length) with satellite geometric factors (cutoff altitude angle, satellite number), and integrates dual frequency GNSS observation data with spatial environment parameters to construct a neural network structure containing physical constraints. The model performance is optimized using an adaptive learning rate and Dropout regularization strategy. The experimental results show that the model can well fit the relationship between spatiotemporal parameters, satellite geometry factors, and ionospheric delay. The root mean square error (RMSE) of the residual value is 0.11 m, the average absolute percentage error (MAPE) of the model is 11.25%, and the accuracy is 88.75%. Meanwhile, the model has a good fitting effect on both the winter data of the Gansu CORS station and the summer data of the Hong Kong CORS Station, indicating that the model has strong adaptability.