Kp指數是用於檢測地磁擾動水平的最重要指標之一，地磁擾動受驅動於太陽風與磁層之間的交互作用。不同情況的太陽風與磁層狀態的耦合使得Kp隱含著多種未知的時變模式。過去的深度學習模型幾乎都使用標準前饋式神經網路預測Kp，然而由於磁暴期間(Kp ≥ 5) 的參數變化劇烈，並且資料相對較少，前饋式神經網路預測磁暴期間Kp的準確度會被局限。本研究使用長短期記憶來預測Kp，長短期記憶是遞迴神經網路的一種變形，能夠儲存隱含在不同輸入特徵之中長期與目標的時間序列相依性。在本研究中，我們設定模型訓練時每次通過當前時步的神經元都包含過去二十四小時的輸入特徵資訊，成功將預測一小時後Kp的準確度比過去研究增加六個百分比。為了滿足不同的需求與條件，本研究也發展了不使用歷史Kp作為輸入參數的模型，並且使用磁層頂日下點對峙距離與磁尾張開程度的參數取代Kp作為模型輸入。結果顯示輸入磁層頂位置參數能有效地降低使用歷史Kp作為模型輸入的依賴性，也能稍微提升磁暴期間的Kp預測準確度。由於磁暴期間較大的地磁擾動會影響地球上的通信設備和電網系統，因此提高Kp的預測準確度有助於太空天氣預報和預警系統的發展，且能有效降低突如其來的極端太空天氣對人類生活的影響。;The Kp index is one of the most important indicators used to monitor the level of the geomagnetic disturbances driven by interactions between the solar wind and the magnetosphere. Different coupling modes of the solar wind-magnetosphere are hidden in time-varying Kp, which makes Kp hard to be forecasted. Most of the previous studies used the general feedforward neural networks (FNN) to predict Kp. However, the accuracy of the FNN is limited to the highly variable and sparse nature in parameters during magnetic storms (Kp ≥ 5). This study uses the long short-term memory (LSTM) to predict the Kp. The LSTM is a variant of the recurrent neural networks (RNN) that can store long-term time series dependencies hidden in different input features. We passed the input feature information for the past twenty-four hours to all neurons during each epoch of model training, successfully increasing the accuracy of predicting Kp for one-hour ahead by 6%. This study also developed a model that did not use historical Kp as an input parameter. We used the parameters of the subsolar distance of the magnetopause and the level of its tail flaring to replace Kp as the model input. Our results show that inputting the magnetopause position parameters can reduce the dependence of using historical Kp as the model input, and the accuracy of a Kp prediction during magnetic storms is slightly increased. Since the large geomagnetic disturbance will affect the communication equipment and power grid systems on the earth, improving the accuracy of the Kp prediction will help develop space weather forecasting and warning systems, which can effectively reduce the impact of extreme space weather on human life.
