Short-term forecast for the occurrence of Large Scale Travelling Ionospheric Disturbances at European middle latitudes using Neural Networks

In this contribution, we propose a new short-term forecast model of Large Scale Travelling Ionospheric Disturbances (LSTIDs) occurrence at specific locations in Europe. The model uses as input data time series with the characteristics of LSTIDs drivers and detected events. The concept underpinning the selection of the input data is based on the phenomenological scenario that the intensity of the auroral electrojets is regulated by the Lorentz force and the Joule heating generates Atmospheric Gravity Waves (AGWs) in the lower thermosphere and LSTIDs in the ionosphere. Based on this scenario, the TEC gradients and the intensity of the auroral electrojets are representative drivers for LSTIDs occurrence. Detected LSTID events and their characteristics are calculated with the HF Interferometry method (HF-INT) over European Digisonde stations. The method looks for coherent oscillation activity in the Maximum Usable Frequency, MUF(3000)F2, and sets bounds to time intervals for which such activity occurs into a given region. HF-INT provides the Spectral Energy Contribution (SEC), which is the contribution of the LSTIDs to the total variability for a given time series. These features (drivers and detected characteristics) are used for the identification of LSTIDs utilizing Machine Learning tools and, more specifically, different types of classifiers, ranging from the traditional k-Nearest Neighbor classifier (k-NN), the Feedforward Neural Networks (FNNs), up to the more advanced Temporal Fusion Transformers (TFTs). Several experiments are performed for two distinct scenarios: (a) values of SEC greater than 50% indicating moderate and strong LSTID activity, and (b) values of SEC greater than 70% indicating strong LSTID activity. The performance is assessed through the F1-score metric, which takes values between 0 and 1 (the higher its value, the better the classifier performance). The forecasting accuracy decreases from 0.9 to 0.6 approximately with increasing forecasting horizon up to two hours ahead for TFT, while the FNNs have the next best performance, and k-NN has inferior performance.