| 摘 要: 为提升流感趋势预测的准确性，针对流感监测序列存在显著非线性、非平稳性以及传统模型在多源驱动因素融合和可解释性方面存在不足的问题，本研究提出一种基于完备集合经验模态分解与自适应噪声(Complete Ensemble Empirical Mode Decomposition with Adaptive Noise，CEEMDAN)和时间融合Transformer(Temporal Fusion Transformer，TFT)的混合预测模型CEEMDAN-TFT。首先利用CEEMDAN对原始ILI序列进行分解，提取不同时间尺度下的本征模态函数(IMF)与趋势项，并采用小波阈值去噪方法对高频噪声分量进行去噪处理，其次将分解后的IMF分量与滞后1周的气象因子对齐融合用于TFT模型的训练，基于美国亚利桑那州2011—2019年流感监测数据进行实验验证，结果表明，CEEMDAN-TFT模型在预测精度上优于CEEMDAN-LSTM等对比模型，在测试集上取得了MAE为0.315、MAPE为8.11%、RMSE为0.395的最优表现，能够为流感趋势预测与公共卫生决策提供技术支撑。 |
| 关键词: 流感预测 完备集合经验模态分解与自适应噪声(CEEMDAN) 时间融合Transformer 多源数据融合 小波去噪 |
|
中图分类号: TP183
文献标识码:
|
| 基金项目: 国家中医药管理局中医药创新团队及人才支持计划项目(N0.ZYYCXTD-D-202208) |
|
| Research on Influenza Prediction Model Based on CEEMDAN-TFT |
|
LIN Defu, CHEN Zhaoxue
|
University of Shanghai for Science and Technology
|
| Abstract: To improve the accuracy of influenza trend forecasting, this study proposes a hybrid prediction model, CEEMDAN-TFT, based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) and the Temporal Fusion Transformer (TFT), in response to the pronounced nonlinearity and non-stationarity of influenza surveillance series, as well as the limitations of traditional models in integrating multi-source driving factors and providing interpretability. First, CEEMDAN was employed to decompose the original influenza-like illness (ILI) series into intrinsic mode functions (IMFs) and a trend component at different time scales, and wavelet threshold denoising was further applied to the high-frequency noise component. Subsequently, the decomposed IMF components were aligned and fused with meteorological factors lagged by one week and then used as inputs for TFT model training. Experiments were conducted using weekly influenza surveillance data from Arizona, USA, spanning 2011 to 2019. The results show that the CEEMDAN-TFT model outperformed comparative models such as CEEMDAN-LSTM in prediction accuracy, achieving the best performance on the test set with an MAE of 0.315, a MAPE of 8.11%, and an RMSE of 0.395. The proposed model can provide effective technical support for influenza trend forecasting and public health decision-making. |
| Keywords: nfluenza forecasting Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN) Temporal Fusion Transformer multi-source data fusion wavelet denoising |
