Ionospheric GPS-VTEC forecasting using hybrid deep learning model (LSTM-CNN)

Abstract

Predictions of the ionosphere are essential for detecting anomalies in space weather conditions. Deep learning technologies applied to ionospheric forecasting have emerged as an exciting field of focus for researchers. This paper presents the forecasting of GPS-VTEC using multilayer perceptron (MLP) and the deep learning model: long short-term memory (LSTM), gated recurrent unit (GRU), and a hybrid model composed of LSTM and a convolutional neural network (CNN). The MLP and the deep learning model are constructed with the GPS-vertical total electron content (GPS-VTEC) time series data estimated over a mid-latitude HUGE Station (${47.834}^{o}N,{7.596}^{o}E$) located in Germany. Factors that influence the variations of GPS-VTEC are identified and used as the input parameters for the deep learning model. The factors are seasonal variation, diurnal variation, magnetic activity, and solar activity. The training is carried out using the data obtained from the years 2010–2014, while the prediction performance of the model is evaluated using the data from the year 2015 (high solar activity). The hybrid LSTM-CNN model performs better than the MLP and the other deep learning model, with a root mean square error (RMSE) of 3.046 total electron content unit (TECU) and a correlation coefficient (R) of 0.896 and is able to capture the diurnal variations of the GPS-VTEC. In addition, the prediction performance of the MLP and the deep learning model is compared with that of IRI-Plas 2017 and NeQuick-2 in capturing the seasonal variation of GPS-VTEC, and it is observed that the IRI-Plas 2017 model provides more accurate predictions than the MLP, the deep learning model, and the NeQuick-2 model during March 2015. However, on average, the deep learning model, the MLP, and the NeQuick-2 model capture the seasonal variations more accurately than the IRI-Plas 2017 at this GPS station.