A Deep Neural Network Model through Automatic Optimization for Ionospheric Electron Density

Ionospheric electron density (N¬_e) is subject to substantial temporal-spatial variations. Specification and forecast these variations have been an important topic in ionosphere research. In this work, we use an automatic optimization for the deep neural network (DNN) models to fit Ne altitude profiles with daily F10.7 index, 3-hourly Ap index, year, month, day of year, magnetic local time (MLT) as the model inputs using long-term ionospheric observations by the Millstone Hill incoherent scatter radar (ISR). To find the best parameters of the DNN model, such as the number of hidden-layers, the number of neurons in each layer and learning rate, etc, the total 23-year data (1995—2017) are used. The mean absolute error (MAE) and relative error (RE) on the data as quantitative measures for the fitting performance show that the automatically optimized DNN model can outperformance the manually-tuned DNN model, where the latter was previously shown significantly better than the single hidden layer NN model. Besides, the annual and diurnal variation patterns shown in electron density observations have been reproduced by the DNN model. Our next step is to develop an altitudinal-resolved DNN model and to investigate their prediction performance. This study will strongly improve the capability of climatologic model to capture meso-scale and dynamic features in the ionospheric data.