MPLinear: Multiscale Patch Linear Model for Long-Term Time Series Forecasting

Qinkai Jiang, Chenyu Hou, Bin Cao, Tianming Zhang, Tiantian Li, Jing Fan

Abstract: Long-term time series forecasting is extensively utilized in practical applications. However, the intricate periodic and trend features inherent in time series pose challenges in exploring their long-term dependency relationships. The current Transformer-based point-wise timing representation and fixed-length patch segmentation inadequately extract various periodic characteristics and local semantic information. Therefore, this paper proposes MPLinear, a Multiscale Patch Linear Model to solve the above problems. We propose a Multiscale Patch strategy to enrich the diverse features of complex periods in time series data and fully extract local deep dependencies as well as global seasonal and trend changes through the Patch Encoding module composed of linear layers. Additionally, to enhance the stability of the model, we perform linear weighted integration of multiscale prediction results by designing an Ensemble Prediction method and further change the training loss to ensure a more robust model. Experimental results show that MPLinear achieves state-of-the-art performance in accuracy and efficiency on seven datasets. The code is available at: https://github.com/Jart39/MPLinear.

External IDs:dblp:conf/iconip/JiangHCZLF24