Neural network-based short-term forecast of Large Scale Travelling Ionospheric Disturbance occurrence above middle and southern Europe

Large-Scale Traveling Ionospheric Disturbances (LSTIDs) are wave-like perturbations of the ambient ionospheric electron density that can have adverse effects on radio wave propagation within or across the ionosphere. Forecast and early identification of LSTIDs is essential information that helps mitigate their impact on radio signal reception. To this end, we develop a new short-term machine learning model to forecast the occurrence of LSTIDs and apply it to specific locations in Europe. The model is trained on time series of LSTID drivers and characteristics of observed LSTIDs. The selection of the input data rests on the assumption that the intensity of the auroral electrojets is regulated by the Lorentz force. The resulting Joule heating generates Atmospheric Gravity Waves (AGWs) in the lower thermosphere and LSTIDs in the ionosphere. The Total Electron Content (TEC) gradients and the intensity of the auroral electrojets are representative drivers of LSTIDs. The characteristics of observed LSTIDs are obtained using the HF Interferometry method (HF-INT) on a network of European Digisonde stations. The method searches for coherent oscillations in the Maximum Usable Frequency for a 3000 km radio path via reflection from the F2 layer (MUF(3000)F2) and finds bounds on time intervals during which such activity occurs in a specific region. HF-INT provides the Spectral Energy Contribution (SEC), which is the contribution of the LSTIDs to the total variability of a given MUF(3000)F2 time series. The LSTID drivers and observed characteristics are leveraged to forecast LSTID occurrences using an advanced Machine Learning method, the Temporal Fusion Transformer (TFT) model. The performance of the TFT model is compared with traditional models, such as the k-Nearest Neighbor classifier (k-NN) and the Feed-forward Neural Networks (FNNs). We consider two distinct scenarios: values of SEC greater than 50% indicating moderate and strong LSTID activity, and values of SEC greater than 70% indicating strong LSTID activity. The performance of the classifiers is assessed through the F1-score metric, which takes values between 0 and 1 (the higher its value, the better the classifier's performance). The forecasting accuracy of TFT decreases from 0.9 to approximately 0.6 as the forecasting horizon increases from 5 min to 2 h ahead. The FNNs have the next best performance, and k-NN has inferior performance. The qualitative analysis of the TFT results provides evidence of a stronger dependence of the performance of the models on historical time series of SEC at the Digisonde locations that are closer to the auroral oval, and much weaker dependence at the lower latitude Digisonde locations. This might be the result of the decreasing LSTID amplitude as it propagates equatorward and may highlight the dissipative nature of the ionospheric medium. Our work demonstrates that a successful forecast of LSTID occurrence requires careful data pre-processing and proper consideration of the drivers, of the propagation pattern, and of other phenomena, such as the damping effect and the interhemispheric propagation.