Data Segmentation based Long-Term Time Series Forecasting

The long-term time series forecasting (LTSF) has ample use cases in aspects of smart city such as energy, finance, and transportation. One challenge of LTSF is capturing temporal correlations that tend to change after the change point, which affects the accuracy of forecasting. This study demonstrates that utilizing linear neural networks for extracting meaning from segmented sub-sequence data can mitigate the influence of change points and enhance the forecasting accuracy of LTSF. Additionally, we introduce a novel method named MSLinear to address the issue of change points that existing linear-centric forecasting models have not tackled. It has multiple heads, and each head evenly partitions time series data and feeds it into linear layers to capture temporal correlations. Test results have shown that this method can increase forecasting accuracy, particularly for datasets with many change points.