Enhancing Solar Driver Forecasting Using Multivariate Transformers: A comprehensive framework to improve Space Weather forecasting.

Resumen

Abstract

We introduce a data-centric framework that forecasts the four key solar drivers, F10.7, S10.7, M10.7, and Y10.7, with PatchTST, a multivariate time-series Transformer that reads an 18-day look-back window and delivers 6-day predictions, capturing short- and long-range patterns in one pass. To offset the natural bias toward quiet Sun conditions, we design a class-balanced loss that weights each sample by its distance from the historical activity distribution, giving storm-level episodes their fair share of influence. Learning is further stabilized with Reversible Instance Normalization, channel-independent embeddings, and an ensemble of weighted MAE and MSE objectives. Tested on the public SET benchmark, the model cuts mean percentage error by 77.7 %and standard percentage deviation by 60.2 % versus SOLAR2000, with gains of more than 80 % during solar-storm periods, and shows less than 2.3 percentage-point drift on a distribution-matched validation set. Training fƒinishes in under eight minutes on a single RTX 3090 and inference stays below two seconds, making the approach viable for near-real-time space-weather forecasting and safer satellite operations.