Learning High-Dimensional Temporal Variations for General Time Series Analysis

Time series analysis underpins a wide range of real-world applications, including weather forecasting, industrial monitoring, and activity recognition. Despite their diversity, these tasks share a central challenge—effectively modeling complex temporal variations. Traditional one-dimensional (1D) approaches often fail to disentangle the intertwined periodic and non-periodic patterns present in real-world data. To address this, we introduce a new paradigm that represents time series through high-dimensional temporal variations, enabling structured modeling of multi-periodic dynamics. Specifically, we decompose temporal signals into intraperiod (high-frequency) and interperiod (low-frequency) variations, transforming 1D sequences into 2D tensors where each dimension captures distinct temporal behaviors. This formulation forms the basis of TimesNet, a unified framework that leverages 2D convolutional kernels to extract rich temporal dependencies and adaptively model multi-periodicity. Beyond 2D, the framework can be extended to 3D temporal variations by incorporating local shapelets as an additional dimension. Extensive experiments show that TimesNet achieves strong performance across forecasting, classification, imputation, and anomaly detection tasks, highlighting the potential of high-dimensional temporal representations for advancing time series analysis.