SCMF: Long-Term Time Series Forecasting Model Based on Series Collapsing and Multiple Periods Fusion

Long-term time series forecasting(LTSF) aims to predict future series based on historical series. Due to the overfitting problem that troubles complex models, lightweight models have become the mainstream trend for LTSF. The recently proposed SparseTSF has excellent performance with parameters below 1K, bringing hope to the research of LTSF. However, SparseTSF has two limitations: (1) The model structure limits its ability to capture complex temporal variations. (2) It can not capture the multiple periodicity of time series, whereas in practice long time series are often interspersed with multiple periods. In order to improve SparseTSF, this paper proposes a LTSF model named as SCMF based on series collapsing and multiple periods fusion. On the basis of SparseTSF, SCMF optimizes the order of learning the intra-period and inter-period relationships, and fuses the dependencies between multiple periods by collapsing series and cross attention. To validate the effectiveness of SCMF, we conducted experiments on seven real-world datasets and the results demonstrate that SCMF is more accurate compared with the state-of-the-art models.