Time Series Decomposition Using Wavelet and Fourier Transforms for Enhanced Solar Flare Forecasting

Abstract

Time series decomposition enables the extraction of
meaningful components by filtering out irrelevant or
noisy frequencies. This study investigates the use of
four transformation techniques—three wavelet-based
(Haar, Symlets, Daubechies) and one Fourier-based
(Discrete Fourier Transform)—to reconstruct multivari
ate time series data for solar flare prediction. The recon
structed data is used to train and evaluate two classifiers:
Time Series Forest (TSF) for time series inputs, and
Random Forest (RF) for transformed non-time-series
representations.
We describe the data preparation pipeline, model train
ing, and comparative performance analysis of TSF
and RF. Additionally, we assess the effectiveness of
data reduction by evaluating model performance on re
duced datasets. Results show that TSF models consis
tently outperform RF models, and that reduced datasets
yield competitive predictive performance. These find
ings suggest the potential for computational efficiency
in near-real-time flare prediction without significant
loss in accuracy. Evaluation metrics include the True
Skill Statistic (TSS) and Heidke Skill Score (HSS2).