電漿泡作為一種電離層不規則體，容易因垂直電漿流破壞高層電漿濃度高、低層電漿濃度低的Rayleigh-Taylor (R-T) 不穩定環境而形成，導致電磁波訊號穿透時發生訊號閃爍。福衛五號上的先進電離層探測儀 (AIP) 提供了距地720公里高的電離層現地量測資料，可有效觀測全球中低緯度的電漿泡並分析垂直電漿流與電漿泡出現之間的關係以及對應的地理位置分布。 本論文將使用AIP從2017年10月31日到2019年10月31日之間量測的電漿資料：經度、緯度、離子濃度標準差、垂直離子流速與實時太陽輻射通量 (F10.7 index) 對類神經網路進行訓練。欲利用類神經網路學習並重現對應環境下電漿泡的發生傾向，旨在建立有能力反映出電漿泡發生潛勢的初步模型，並利用網路模型反映出的預測能力來探討資料處理流程與網路模型的定義取捨。 最終訓練出的預測模型中，二月、十月與十二月的電漿泡發生潛勢預測模型的地理分布預測表現最理想。其中，十二月的預測模型表現更為優異，甚至可以取代十一月與一月的預測模型。根據訓練的結果可對十月至二月的全球中低緯度的電漿泡發生潛勢提供具有參考性的預測。本篇研究結果亦會呈現處理過後的資料特徵及在本研究中利用泛化能力的表現來改善電漿泡預測模型的過程。 ;Equatorial plasma bubbles (EPBs) usually cause stronger signal scintillation in space-earth communications compared to other ionospheric irregularities. The most commonly accepted theory for the EPBs formation mechanism is that they are developed from a Rayleigh-Taylor instability (R-T instability) . But to date, predicting the likelihood EPBs’ appearance is still challenging, despite scientists academics already having a basic consensus about the formation mechanism of EPBs. The Advanced Ionospheric Probe (AIP) is a science payload installed on FORMOSAT-5, providing high resolution in-situ plasma measurement at altitude 720 km. According to the factors contributing to the EPBs’ triggering process, this research collected vertical ion drift velocity, ion number density, the corresponding latitude and longitude from AIP, attached with real-time F10.7 index. After collecting and filtering the data, an empirical model was built to predict the potential appearance of EPBs via Artificial Neural Network (ANN) . Among the prediction models finally trained, the prediction models for the occurrence potential of plasma bubbles in February, October and December performed best in terms of geographic distribution. Among them, the forecast model for December performed even better and even replaced the forecast models for November and January. According to the training results, it can provide a reference prediction for the occurrence potential of plasma bubbles in the middle and low latitudes of the world from October to February. The results of this study will also present the characteristics of the processed data and the process of using the performance of the generalization ability to improve the plasma bubble prediction model in this study.
