TSINet: A temporal-channel factorized mixing and spectral enhanced interactive network for time series forecasting

Time series forecasting is crucial in fields like energy, meteorology, and power systems, where accurately modeling both long-term and short-term dependencies is vital. While Transformer-based models are effective at capturing long-term patterns through self-attention, their performance is constrained by sensitivity to high-frequency noise and limited use of spectral information. To tackle these issues, this work presents a temporal-channel factorized mixing and spectral enhanced interactive network, named TSINet. Specifically, TSINet employs a Time-channel factorized mixing module, which uses a factorization strategy to facilitate cross-dimensional interactions between time steps and channels, thereby reducing redundant noise; a Spectral information enhanced decomposition mixing module to enhance high-frequency spectral information and improve the extraction of key signal features; and an Interactive representation shared-uniqueness prediction module that combines large and small convolutional kernels to jointly capture global trends and local variations. Through tailored fusion and separation strategies, TSINet effectively models the multi-level structure of time series data. Experimental outcomes reveal that TSINet consistently delivers superior forecasting results compared to leading models across seven real-world datasets in areas such as electricity, weather, and traffic.