TF-CL: Time Series Forcasting Based on Time-Frequency Domain Contrastive Learning

Wen Li, Yun Gu, Shouguo Du

Abstract: Accurately forecasting complex time series data is an essential task across a wide range of application scenarios. While the mainstream research on time series forecasting based on contrastive learning currently focuses on time-domain data augmentation to calculate contrastive loss, the importance of frequency domain analysis is often overlooked. Frequency domain analysis can reveal the main frequency components of the data, providing a unique perspective for understanding the periodicity, oscillation, and fluctuation characteristics of the data. We proposes time series forcasting based on time-frequency domain contrastive learning (TF-CL). TF-CL is constructed based on the Bayesian time series hypothesis and includes three core components: an encoder, a trend feature extractor, and a seasonal feature extractor, each utilizing contrastive loss functions in the time and frequency domains for feature extraction. Empirical analysis conducted on six cross-domain datasets shows that, compared to current benchmark methods, our model demonstrates superior predictive capabilities in the majority of cases. Furthermore, through detailed ablation studies, TF-CL further validates the critical role of integrating frequency domain data enhancement techniques with time-frequency domain analysis in improving predictive performance.