Abstract: To address inadequate adaptability of existing time series forecasting methods under rapidly changing data distributions and difficulty in balancing prediction accuracy with computational overhead, a time series forecasting model based on dynamic weighted ensemble(TFEM) was proposed. Firstly, in the time domain module, a Low-Rank Self-Attention mechanism(LRSA)was designed by computing attention through projecting high-dimensional features into a low-dimensional space, which reduced complexity while maintaining long-range dependency modeling. Simultaneously, in the frequency domain module, the signal was decomposed into dominant frequency components and non-stationary residuals to model global trends and local abrupt changes separately, enhancing the modeling capability for complex time series. Finally, at the ensemble level, a long-short term harmonic weighting mechanism was proposed, where long-term weights captured global trends robustly through recursive updating, short-term weights responded promptly to data distribution shifts via a multilayer perceptron, and a smoothing factor was incorporated to suppress severe fluctuations in weights. Experimental results demonstrate that compared with the online ensemble method OneNet(Online Network), TFEM reduces the Mean Squared Error (MSE) by 6.4%~44.8% and the Mean Absolute Error (MAE) by 2.8%~26.8% across seven benchmark datasets, while also reducing the parameter count by 69.4% and inference time by 50.5% on the ETTh1 dataset. The experimental findings show that TFEM enhances prediction accuracy while reducing computational overhead, providing a feasible solution for time series forecasting in resource-constrained scenarios.

Key words: time series, ensemble learning, time domain, frequency domain, dynamic weighting

摘要： 针对现有时间序列预测方法在数据分布快速变化时适应性不足且难以平衡预测精度与时空开销的问题，提出一种基于动态加权集成的时序预测模型(TFEM)。首先，在时域模块中设计低秩自注意力机制(LRSA)，通过将高维特征投影至低维空间计算注意力，在降低复杂度的同时保持长程依赖建模；同时，在频域模块中将信号分解为主导频率分量和非平稳残差，分别建模全局趋势与局部突变，提升对复杂时序的建模能力；最后，在集成层面提出长短期谐衡加权机制，其中长期权重通过递归更新稳健捕捉全局趋势，短期权重借助多层感知机及时响应数据分布的突变，并结合平滑因子抑制权重的剧烈波动。实验结果表明，与在线集成的 OneNet(Online Network)相比，TFEM 在 7 个基准数据集上的均方误差(MSE)降低了 6.4%~44.8%，平均绝对误差(MAE)降低了 2.8%~26.8%，且在 ETTh1 数据集上参数量减少了 69.4%，推理时间减少了 50.5%。实验结果表明，TFEM 在提升预测精度的同时能够降低计算开销，为资源受限场景下的时序预测提供了可行方案。

关键词: 时间序列, 集成学习, 时域, 频域, 动态加权