LFformer: Line Projection Decomposed Transformer for Long-Term Series Forecasting

Applications for financial risk assessment as well as weather forecasting are greatly dependent on the time period of the projection to be projected in time series. Although they are capable of taking on the most important role from the long run by the use of transformer-based models because of their self-lead mechanism, they also have difficulty with complicated temporal patterns because of the increased computational difficulty in this regard. We suggest using LFformer, a novel method that increases both time series forecasting efficiency and correctness to get around this drawbacks. Through the use of specific attention head optimization and linear projections to reduce time complexity, LFformer guarantees computing scalability. We propose an efficient and flexible way to replace the conventional decomposition approaches of Fourier Seasonal Analysis (Fourth Generation) as we attempt to address the complicated temporal trends that are part of the data we use today. According to the results of our experiment, the LFformer continuously performs well in many datasets, demonstrating how well it might be used with a robust time series forecasting tool.