An influenza forecasting model based on RG-EEMD-TFT

Abstract

Objective: To improve the accuracy of influenza forecasting, this study proposes a Reconstruction Grouping–Ensemble Empirical Mode Decomposition–Temporal Fusion Transformer (RG-EEMD-TFT) model designed to address the nonlinear, non-stationary, and multi-scale characteristics of influenza surveillance time series, while addressing the limitations of conventional deep learning models in terms of feature redundancy and interpretability. Methods: Weekly influenza-like illness (ILI) data from Colorado, USA, were used in this study. Ensemble Empirical Mode Decomposition (EEMD) was first applied to decompose the original series into several intrinsic mode functions and a residual component. A reconstruction grouping (RG) strategy was then used to reorganize the decomposed signals into high-frequency, seasonal, and trend components, which were subsequently used as inputs to a Temporal Fusion Transformer (TFT). Hyperparameter optimization was performed using the Neural Network Intelligence (NNI) framework. Results: Models incorporating signal decomposition consistently outperformed their non-decomposed counterparts. The proposed RG-EEMD-TFT model outperformed all comparison models, achieving a mean absolute error (MAE) of 0.318, a mean absolute percentage error (MAPE) of 8.012%, and a root mean square error (RMSE) of 0.409. Conclusion: By integrating EEMD, reconstruction grouping, and TFT, the proposed RG-EEMD-TFT model effectively captures the multi-scale temporal characteristics of influenza activity and improves forecasting accuracy.