A Fast Season Length Estimation Using Sliding Discrete Fourier Transform for Time Series Streaming Data
Thanapol Phungtua-Eng, Yoshitaka Yamamoto · 2024
Season length estimation (SLE) in time series data is critical for identifying cycle lengths in time series data, especially in streaming environments where data is continuously updated. Traditional SLE methods, designed for offline mode, require the entire dataset, which is challenging for real-time analysis. Adapting these methods to online mode often involves a sliding window technique, but they still incur a computational cost of$\mathrm{O}(N\log N)$per timestamp, where$N$is the length of the sliding window. This computational cost makes them less feasible for online streaming data applications. To address this limitation, we present OnlineSLE, an online season length estimation method designed for time series streaming data. OnlineSLE utilizes the sliding discrete Fourier transform (SDFT) to reduce the computational cost to$\mathrm{O}(N)$for each timestamp, which allows for faster processing. Our extensive evaluation of various synthetic and real-world datasets demonstrates that OnlineSLE not only accelerates the computational process, but also maintains high accuracy in season length estimation. The results underscore the significant advances our method achieves in online season length estimation for time series streaming data.