DFCon: Dominant frequency enhanced ultra-long time series contrastive forecasting

Ultra-long time series forecasting (ULTSF) is crucial for fields like energy management, traffic planning, and climate prediction. However, as the forecast horizon increases, concept drift becomes a major challenge, as a fixed-length historical window struggles to generalize ultra-long temporal patterns. Extending the input series length increases computational costs and demands a higher model capacity for capturing longer temporal dependencies. To address these issues, we propose DFCon, a dominant frequency enhanced contrastive learning framework for ULTSF. DFCon combines dilated convolutions for feature extraction and multi-layer perceptrons for forecasting, with a dual contrastive loss based on dominant frequency enhancement. We introduce Temporal DFCon, which enhances the model's sensitivity to these frequency-domain features during training, thereby improving its ability to model global temporal dependencies in the input series. Furthermore, cross-window Autocorrelated DFCon is proposed, which mitigates concept drift by constructing autocorrelated relative positive and negative samples without introducing noisy data. Experiments on five benchmark datasets show that DFCon outperforms existing methods, demonstrating its effectiveness in ULTSF. The code for this work is publicly available at: https://github.com/coding4qq/DFCon.