T-FORS LSTID Forecasting with TFT

Short-Term High-Accuracy (ST-HA) model, aimed at forecasting the LSTID occurrence with a horizon of a few hours (1 to 3 hours). The approach followed the auroral electrojet indices IL and IU provided by the FMI IMAGE network, the Gradient GNSS TEC Activity Index provided by DLR and the archived estimates of the Spectral Energy Contribution (SEC) calculated with the HF-INF method based of the analysis of MUF characteristics over Digisonde stations at European middle latitudes. A Temporal Fusion Transformer classifier is then trained to forecast the SEC of detected LSTIDs over Digisonde stations.

The main objective of the modelling is to forecast the occurrence and propagation characteristics of LSTIDs. We concentrated on the Short-Term High-Accuracy (ST-HA) models, aimed at forecasting the LSTID occurrence with a horizon of a few hours (1 to 3 hours). This approach casts the forecasting task as a binary classification problem using multivariate time series data. The input data consists of: (a) Auroral electrojet indicators (IL and IU) from the Scandinavian longitudes (provided by the FMI IMAGE network). (b) Gradient GNSS TEC Activity Index (provided by DLR). (c) Spectral Energy Contribution (SEC) values derived with the HF Interferometry method at mid-latitude Digisonde locations (the HFI calculations are provided by the Ebro Observatory). The output is a binary classification indicating whether a Large-Scale Travelling Ionospheric Disturbance (LSTID) event will occur within the next two hours. An event is considered to have occurred if any forecasted SEC value exceeds a predetermined threshold.