Low Carbon Considerations in Space Weather AI: Weighted Ensemble Feature Selection for FoF2 Forecasting

Abstract

The increasing computational cost of machine learning models impacts carbon emissions. A more complex model produces more $\mathbf{C O}_{\mathbf{2}}$ and requires longer training times. This poses a significant challenge in predicting the critical frequency of the F2 layer (foF2), a crucial parameter for HF radio communications, which increasingly relies on complex deep learning models such as Transformer-based models that improve accuracy but burden the computational budget. This paper proposes Green-Aware Weighted Ensemble Feature Selection (GAWEFS), a framework that addresses this problem at the data preprocessing stage. Rather than redesigning model architectures, GAWEFS selects which features to feed them. It combines filter, wrapper, and embedded selection techniques, samples data through a temporal window strategy, and assigns weights by voting on three criteria, including prediction error, carbon emissions, and training time. Tested on the Informer model using hourly low-latitude ionosonde data from 2006 to 2018, GAWEFS significantly mitigates carbon emissions and reduces training time while maintaining prediction accuracy comparable to full-training baselines. Specifically, GAWEFS reduced training time by approximately 86% and carbon emissions by approximately 85% compared to the best-performing single feature selection method (Boruta). Furthermore, it achieved an Accuracy-per-Gram score of 0.263, which is 4.4 times higher than the Full Features baseline and nearly 7 times higher than Boruta, with only a modest 8.5% increase in RMSE. The carbon investment of GAWEFS was offset within 3.1 retraining cycles, yielding projected savings of $\mathbf{1 2 7. 3} \mathbf{~ g C O}_{\mathbf{2}}$ by the 14th cycle. For systems requiring continuous retraining, these savings are compounded.