A Long-term Time Series Forecasting method with Multiple Decomposition

In various real-world applications such as weather forecasting, energy consumption planning, and traffic flow prediction, time serves as a critical variable. These applications can be collectively referred to as time-series prediction problems. Despite recent advancements with Transformer-based solutions yielding improved results, these solutions often struggle to capture the semantic dependencies in time-series data, resulting predominantly in temporal dependencies. This shortfall often hinders their ability to effectively capture long-term series patterns. In this research, we apply time-series decomposition to address this issue of long-term series forecasting. Our method involves implementing a time-series forecasting approach with deep series decomposition, which further decomposes the long-term trend components generated after the initial decomposition. This technique significantly enhances the forecasting accuracy of the model. For long-term time-series forecasting (LTSF), our proposed method exhibits commendable prediction accuracy on four publicly available datasets—Weather, Electricity, Traffic, ILI—when compared to prevailing methods. The code for our method is accessible at https://github.com/wangyang970508/LSTF_MD.