長短期記憶深度學習類神經網路於預測電離層垂直總電子含量之應用

The Application of LSTM Deep Learning Artificial Neural Network on the Prediction of the Ionospheric Vertical Total Electron Content

田凱元(Kan-Yuan Tian)

指導教授 ： 王立昇

國立臺灣大學/工學院/應用力學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202104528

摘要

隨著行動裝置的普及與科技的快速發展，使用者對於全球衛星定位系統(Global Positioning System, GPS)之定位精度要求也日趨提升，對於目前使用普及之單頻衛星訊號接收機而言，電離層誤差為影響定位效果最重要的誤差項之一，故使用準確且即時的垂直總電子含量(vertical total electron content, VTEC)計算電離層誤差尤為重要。
本研究使用類神經網路預測的方法以期求得準確之VTEC數值，首先試驗了陽明山地區與北京地區之VTEC、太陽黑子數、F10.7參數、ap指數、Dst指數以及行星際磁場南向分量等組成之較佳輸入參數組合，發現具不同電離層結構之兩地區其較佳輸入參數組合亦會不盡相同。本研究亦比較了單向LSTM最後輸出模型、單向LSTM序列輸出模型、雙向LSTM最後輸出模型、雙向LSTM序列輸出模型以及BPNN模型之VTEC預測效能，發現雙向LSTM序列輸出模型具有最佳之預測表現，最後將陽明山地區和北京地區之雙向LSTM序列輸出模型與傳統BPNN模型、C1PG一日預測電離層地圖、IRI-2016模型進行預測效能比較，並使用CODG電離層地圖當作參考標準，於2016年測試資料發現雙向LSTM序列輸出模型之預測表現最佳。
為檢驗預測結果之效能，本研究之定位實驗先計算出陽明山測站天頂穿刺點位置類神經網路預測值與C1PG於相同穿刺點位置預測值之偏差值，再於C1PG電離層地圖各衛星穿刺點位置同加上此一偏差值，代入定位演算法進行定位後，於平面定位有明顯的精度改善效果。

關鍵字

長短期記憶 ； 類神經網路 ； 電離層 ； 垂直總電子含量 ； 單頻單點定位

並列摘要

Along with the broad usages of mobile devices and the rapid development of technologies, the requirements for the positioning accuracy of the Global Positioning System (GPS) become much more demanding. For the currently popular single-frequency single-point GPS receivers, the ionospheric delay is one of the most important errors affecting the positioning performance. Therefore, it is particularly critical to have accurate and real-time vertical total electron content (VTEC) data to calculate the ionospheric error.
In this study, the artificial neural network (ANN) methods is used to predict accurate VTEC values. Some closely related data, such as the VTEC, sunspot number, F10.7 parameter, ap index, Dst index, and the IMF-Bz in the Yangmingshan area and Beijing area are first obtained from various sources. These data are then fed into the artificial neural network algorithm to train the parameters in the algorithm. It is found that the optimal sets of parameters for the two regions which have different ionospheric structures are also different. Moreover, five different schemes of ANN, including the LSTM last output model, the LSTM sequence output model, the Bi-LSTM last output model, the Bi-LSTM sequence output model, and the BPNN model, are compared for their performance of prediction. It is found that the Bi-LSTM sequence output model outperform the other models. To compare the performance of the ANN algorithm with other prediction methods, the Bi-LSTM sequence output model, the traditional BPNN model, the C1PG one-day prediction ionospheric maps, and the IRI-2016 model are used to predict the VTEC in the Yangmingshan area and the Beijing area, with the CODG ionospheric maps being used as reference standards. It is found that, for the test date in 2016, the Bi-LSTM sequence output model has the best prediction performance.
To assess the efficiency of the ANN prediction algorithm, the predicted VTEC is used to compute the ionospheric error at the position of the zenith pierce point at the Yangmingshan station. The difference between the predicted VTEC and the VTEC　data from C1PG is found and used to compensate the VTEC data at the neighboring pierce points on the C1PG ionospheric map. The revised ionospheric map is then adopted to perform positioning experiment at the Yangmingshan station. It is observed that there is a high probability of reduction of the positioning error on horizontal plane. This result shows that developed ANN algorithm is indeed helpful in the enhancement of the performance of single-frequence GPS receivers.

並列關鍵字

long short-term memory ； artificial neural network ； ionosphere ； vertical total electron content ； single-frequency single-point positioning

參考文獻

[1]John A. Klobuchar. Ionospheric time-delay algorithm for single-frequency GPS users. IEEE Transactions on Aerospace and Electronic Systems, 1987, 3:325-331.

[2]Sigrid K. Llewellyn, Rodney B. Bent. DOCUMENTATION AND DESCRIPTION OF THE BENT IONOSPHERIC MODEL. Indialantic, Florida: Atlantic Science Corporation, 1973.

[3]http://irimodel.org/

[4]Karthik.P, D. Venkata Ratnam, Nagireddy Vanga, P. S. Brahmanadam, B. Sada Siva Rao and K. Sarat Kumar. Auto Regressive Ionospheric Prediction model for GPS Applications. International Journal of Computer Applications, 2012, vol. 48, no. 4.

[5]G. Sivavaraprasad, V. S. Deepika, D. SreenivasaRao, M. Ravi Kumar, M. Sridhar. Performance evaluation of neural network TEC forecasting models over equatorial low-latitude Indian GNSS station. Geodesy and Geodynamics, 2020, 11:192-201.