Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FPN-fusion: Enhanced Linear Complexity Time Series Forecasting Model

Chu Li, Pingjia Xiao, Qiping Yuan

TL;DR

This work tackles scalable, accurate time series forecasting by introducing FPN-fusion, a linear-time model that abandons traditional trend/season decomposition in favor of a Feature Pyramid Network (FPN) to extract multi-scale temporal features. The method pairs FPN-based feature extraction (with deep-trend and shallow-seasonal cues) with a UNet-like multi-level fusion to integrate information across scales, yielding two lightweight variants (FPNLinear and FPNMLinear) and the full FPN-fusion architecture. Empirically, FPN-fusion achieves substantial accuracy gains over DLiner (average MSE reduction of 16.8% and MAE reduction of 11.8% across 32 tests) and competitive results against PatchTST, all while maintaining $O(L)$ computational complexity and significantly lower compute costs than transformer-based baselines. This approach offers a practical, scalable solution for real-world forecasting tasks across univariate and multivariate settings, with potential extensions to multi-step prediction and online learning.

Abstract

This study presents a novel time series prediction model, FPN-fusion, designed with linear computational complexity, demonstrating superior predictive performance compared to DLiner without increasing parameter count or computational demands. Our model introduces two key innovations: first, a Feature Pyramid Network (FPN) is employed to effectively capture time series data characteristics, bypassing the traditional decomposition into trend and seasonal components. Second, a multi-level fusion structure is developed to integrate deep and shallow features seamlessly. Empirically, FPN-fusion outperforms DLiner in 31 out of 32 test cases on eight open-source datasets, with an average reduction of 16.8% in mean squared error (MSE) and 11.8% in mean absolute error (MAE). Additionally, compared to the transformer-based PatchTST, FPN-fusion achieves 10 best MSE and 15 best MAE results, using only 8% of PatchTST's total computational load in the 32 test projects.

FPN-fusion: Enhanced Linear Complexity Time Series Forecasting Model

TL;DR

This work tackles scalable, accurate time series forecasting by introducing FPN-fusion, a linear-time model that abandons traditional trend/season decomposition in favor of a Feature Pyramid Network (FPN) to extract multi-scale temporal features. The method pairs FPN-based feature extraction (with deep-trend and shallow-seasonal cues) with a UNet-like multi-level fusion to integrate information across scales, yielding two lightweight variants (FPNLinear and FPNMLinear) and the full FPN-fusion architecture. Empirically, FPN-fusion achieves substantial accuracy gains over DLiner (average MSE reduction of 16.8% and MAE reduction of 11.8% across 32 tests) and competitive results against PatchTST, all while maintaining computational complexity and significantly lower compute costs than transformer-based baselines. This approach offers a practical, scalable solution for real-world forecasting tasks across univariate and multivariate settings, with potential extensions to multi-step prediction and online learning.

Abstract

This study presents a novel time series prediction model, FPN-fusion, designed with linear computational complexity, demonstrating superior predictive performance compared to DLiner without increasing parameter count or computational demands. Our model introduces two key innovations: first, a Feature Pyramid Network (FPN) is employed to effectively capture time series data characteristics, bypassing the traditional decomposition into trend and seasonal components. Second, a multi-level fusion structure is developed to integrate deep and shallow features seamlessly. Empirically, FPN-fusion outperforms DLiner in 31 out of 32 test cases on eight open-source datasets, with an average reduction of 16.8% in mean squared error (MSE) and 11.8% in mean absolute error (MAE). Additionally, compared to the transformer-based PatchTST, FPN-fusion achieves 10 best MSE and 15 best MAE results, using only 8% of PatchTST's total computational load in the 32 test projects.
Paper Structure (17 sections, 5 equations, 2 figures, 5 tables)

This paper contains 17 sections, 5 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Linear Time Series Models
  4. Transformer-Based Models
  5. Time Series Decomposition Techniques
  6. Proposed Method
  7. FPN of time series
  8. FPNLinear and FPNMLinear
  9. FPN-fusion model
  10. Evaluation metric
  11. Experiments
  12. Platform
  13. Compared SOTA methods
  14. Ablation Study
  15. model efficiency
  16. ...and 2 more sections

Figures (2)

  • Figure 1: The pipeline of existing Model TSF solutions.
  • Figure 2: (a) Feature Pyramid Network for time series analysis. (b) FPNLinear algorithm structure. (c) FPNMLinear algorithm structure. (d) FPN-fusion algorithm structure.