<p>Rockburst data’s high-dimensional and imbalanced nature makes traditional numerical simulations and machine learning methods struggle with prediction accuracy and capturing temporal dependencies. Deep time-series models typically rely on a large number of parameters, are less effective in predicting minority classes, and are prone to overfitting. Frequency-domain analysis, which transforms time-series data into the frequency domain, offers an effective alternative for analysis and prediction. This paper presents the Fourier-based Kolmogorov-Arnold Network (FKAN), which applies the Fourier transforms better to capture periodicity and temporal dependencies in time-series data. The model retains the low-parameter, high-performance benefits of the Kolmogorov-Arnold Network (KAN) while addressing its limitations in time-series prediction. Experimental comparisons with the classical and state-of-the-art predictive models show that FKAN significantly outperforms these baseline models with fewer parameters. Additionally, a slight increase in parameters allows FKAN to achieve entirely accurate predictions in minority class tasks, surpassing traditional high-parameter models. Focusing on the complex task of rockburst prediction, this study highlights FKAN’s superior low-parameter, high-performance capabilities and its strong performance in minority class predictions. The combination of frequency-domain analysis with KAN demonstrates considerable potential for time-series data processing, offering a novel approach to frequency-based information processing and advancing time-series prediction models.</p>