CLEARflare-LSTM

A Deep Long Short-Term Memory (LSTM) framework for solar flare forecasting using Spaceweather HMI Active Region Patch (SHARP) parameters. The CLEARflare-LSTM model predicts the probability of M-class and above (M+) flares for each HARP active region and for the full solar disk over the next 24 hours, based on HMI magnetogram data from the preceding 24 hours. The model is trained using the GOES science-quality flare summary dataset, combined with operational flare summary data and science-quality flare location data.

The full disk probabilities are calculated from the active region probabilities by adopting the maximum probabilities among all the HARPs.

The forecast is fully generated by the trained Logistic Regression models.

During the training process, we associate HARP regions with NOAA Active Regions using metadata provided in the SHARP FITS header keywords.

We compute the standard deviation from the outputs of 30 bootstrapped model runs. The upper and lower bounds are defined as the mean ± one standard deviation, approximately corresponding to the 16th and 84th percentiles.

The full disk forecast is valid for +- 68 degrees from Sum central meridian.

We do not currently provide calibration levels.

The model forecast "M and above" probabilities. (with uncertainties) for the next 24 hours from the issue time.