ASTD: Automatic Seasonal-Trend Decomposition for Time Series
Bowen Chen · Proceedings/Proceedings of the ... International Conference on Software Engineering and Knowledge Engineering · 2023
The rapid and accurate decomposition of multiperiod time series is crucial for reliable forecasting, anomaly detection, and classification.However, the traditional approach of first detecting the periodicity and then selecting from a range of decomposition algorithms based on the periodicity results in inefficiencies and complexity.To address this challenge, we propose the automatic seasonal-trend decomposition (ASTD), a unified method for automatic time series decomposition.With the ASTD, users no longer need to worry about whether their time series is multi-period, single-period, or aperiodic.They simply provide the time series, and the ASTD automatically returns the final decomposition results.Careful consideration of runtime cost and accuracy requirements has been taken in the design of the ASTD, which has an overall time complexity of O(N logN ).Extensive experimental results show that the proposed ASTD outperforms other state-of-the-art decomposition algorithms in terms of minimum mean square error (MSE) and mean absolute error (MAE).Notably, when applied to the Taylor dataset, the ASTD is approximately 3 times faster than other baseline decomposition algorithms.