Benchmarking Deep Learning Models for Real‐Time Prediction of Post‐Peak Solar Flare Irradiance Using FISM2 Data

Abstract

This study presents a deep learning framework for real‐time prediction of post‐peak solar flare irradiance. The model targets the soft X‐ray band (0.1–0.8 nm) and three extreme ultraviolet (EUV) wavelength bands (9.0–9.9 nm, 13.0–13.9 nm, and 30.0–30.9 nm), and produces forecasts at 1‐min cadence for 3 hr after the flare peak. We used 964 M‐ and X‐class flare events from January 2003 to March 2023, with 60 min of pre‐peak Geostationary Operational Environmental Satellite (GOES) X‐ray flux as input and Flare Irradiance Spectral Model Version 2 (FISM2) reconstructed irradiance as the reference target. We trained and evaluated multiple deep learning architectures, including fully connected neural network (FCN), long short‐term memory (LSTM), gated recurrent unit (GRU), recurrent neural network (RNN), and Transformer models, and compared them with statistical baseline models. The FCN model achieved correlation coefficients of 0.88 in the 0.1–0.8 nm X‐ray band and 0.55, 0.82, and 0.66 in the 9.0–9.9 nm, 13.0–13.9 nm, and 30.0–30.9 nm EUV bands, respectively. The GRU model achieved the highest mean correlation coefficient among the tested deep learning models, while random forest regression provided the strongest statistical baseline model. Nevertheless, the FCN model remained competitive in error‐based metrics and showed particularly strong absolute peak error performance in the soft X‐ray band and competitive error‐based performance in the 30.0–30.9 nm band. These results demonstrate the feasibility of real‐time post‐peak irradiance forecasting across multiple X‐ray and EUV bands and provide a benchmarked foundation for future full‐spectrum irradiance prediction to support ionospheric modeling and space weather operations.