Research On Ionospheric Parameter Prediction Model Based On Improved LSTM

Artificial intelligence technology is an emerging research direction in the study of ionospheric parameter modeling.The transmission signal of the navigation satellite positioning system is directly affected by the change of the ionosphere.Therefore,the prediction of ionospheric parameters is of great significance in the correction of navigation and positioning delay.This work investigates the research background,significance and main research methods of this topic,and found that traditional forecasting methods mainly use Kalman filter forecasting,EMD decomposition method,artificial neural network ANN method,etc.These methods model the ionospheric parameter forecasting as a time series for forecasting problems,these methods work well in forecasting periodic trend components,but the forecasting accuracy is generally not high in the detailed part of the sequence.To address these drawbacks of existing models,this paper proposes the Gated-AttentionBiLSTM model.The model includes a bidirectional LSTM coding layer,a global attention layer,a gating mechanism layer,and an output layer.First of all,the main predicted ionospheric parameter TEC is in this paper,which are encoded by a bidirectional LSTM network to obtain an embedded representation of each position on the TEC parameters trajectory.A global attention mechanism is then done on the encoded output,and adaptive weighting is done on each location point on the sequence to obtain the trajectory representation of the TEC parameters.Afterwards,the encoded features are adaptively learned by gating mechanism to learn the importance of each feature.With the combination of attention mechanism and gating mechanism,the Gated-Attention-LSTM model has higher prediction accuracy and can better take into account both periodic and non-periodic parts of the time series.Furthermore,it is more capable of learning the intrinsic patterns in the detailed prediction of time series,which has the characteristics of small prediction error and easy engineering implementation.Comparing from the experimental indexes,the improved model improves 85.85%compared to Kalman filter MSE,62.38%compared to RMSE,and 3.8%compared to MAPE absolute;47.55%compared to EMD decomposition+LSTM,27.58%compared to RMSE,and 1.2%compared to MAPE absolute.The experiment demonstrates that the four evaluation indexes Gated-Attention-LSTM proposed in this paper improves the prediction accuracy significantly than the traditional method Kalman filter,EMD+LSTM model effect.This proves that the proposed model in this paper is effective and has some practical significance in navigation positioning delay correction.