Enhancing long-term traffic prediction through hybrid TimesNet-informer architecture

This paper presents MixModel, a novel hybrid deep learning architecture that integrates TimesNet’s frequency-aware convolutional feature extraction with Informer’s long-range dependency modeling. Unlike existing single-model approaches, MixModel employs parallel processing and attention-weighted fusion to simultaneously capture both periodic patterns and global temporal structures. A dedicated fusion mechanism aligns features across branches, enabling enhanced accuracy and robustness in long-term time series prediction. Experimental results confirm that the proposed architecture significantly improves predictive performance over state-of-the-art baselines while maintaining efficient computation.