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

Time series decomposition enables the extraction ofmeaningful components by filtering out irrelevant ornoisy frequencies. This study investigates the use offour transformation techniques—three wavelet-based(Haar, Symlets, Daubechies) and one Fourier-based(Discrete Fourier Transform)—to reconstruct multivariate time series data for solar flare prediction. The reconstructed data is used to train and evaluate two classifiers:Time Series Forest (TSF) for time series inputs, andRandom Forest (RF) for transformed non-time-seriesrepresentations.We describe the data preparation pipeline, model training, and comparative performance analysis of TSFand RF. Additionally, we assess the effectiveness ofdata reduction by evaluating model performance on reduced datasets. Results show that TSF models consistently outperform RF models, and that reduced datasetsyield competitive predictive performance. These findings suggest the potential for computational efficiencyin near-real-time flare prediction without significantloss in accuracy. Evaluation metrics include the TrueSkill Statistic (TSS) and Heidke Skill Score (HSS2).