[P-EM12] Coupling Processes in the Atmosphere-Ionosphere System

[PEM12-P36] BP Neural Network Model for Ionospheric TEC Prediction: Incorporating the Impact of Geomagnetic Field

*ZIFAN XU¹、GUANYI MA¹ (1.The National Astronomical Observatories of the Chinese Academy of Sciences)

キーワード：Total Electron Content (TEC) , Ionosphere, Electrodynamic , BP Neural Network

The Total Electron Content (TEC) variation in the ionosphere represents a continually evolving and intricate electrodynamic process that is shaped by diverse factors, including solar radiation, geomagnetic filed, and so on. We aim to develop an ionospheric TEC model based on BP neural network, to describe this complex process of change by using TEC data from the Global Ionospheric Map (GIM) released by the International GNSS Services (IGS). The BP neural network model consists of an input layer, a hidden layer, and an output layer, which are connected through a feedforward network using the gradient descent backpropagation algorithm. During the experimental phase, the input layer includes time parameters (Day of year, Local time) and solar parameters (F10.7, Lyman alpha), the hidden layer have 20 nodes, and the output layer is TEC value. We use the TEC data provided by the IGS GIM product from 2008 and divide it into a training set (70%), a validation set (15%), and a test set (15%). The experimental results demonstrate that there are issues of mismatch or overfitting between the validation and test data. And the data performance for training, validation, and testing exhibits a high degree of consistency. The RMSE value stabilize at around 1.0 when the neural network converges, and the TEC accuracy range from -7 to 4 TECU. Taking into account the impact of geomagnetic field on the ionosphere, we add the Dst-index (geomagnetic index) into the input layer of the model. This adjustment significantly improves the model accuracy. Specifically, after adding the Dst-index, the RMSE value of the neural network further decrease to approximately 0.8, and the TEC accuracy improve to within the range of -5 to 3 TECU. Basing on these experimental results, we can tentatively conclude that the introduction of the Dst-index parameter effectively enhances the accuracy of model made by the neural network.