Abstract: To address high computational cost and limited structural adaptability of Transformer-based methods for long-term time series forecasting, this paper proposes FiM-Transformer, a framework integrating frequency-domain priors with linear state space models. It first reconstructs multivariate historical sequences into variable-oriented tokens through an inverted-dimension strategy, and then employs a multi-view super tokenization mechanism to jointly model temporal, frequency-spectrum, and evolutionary features with adaptive gating. In addition, a flexible backbone network based on structural risk minimization and a dynamic trend decomposition strategy are introduced to achieve heterogeneous modeling of trend, seasonal, and residual components while adaptively adjusting model capacity. Experimental results on Electricity Transformer Temperature, Electricity Consuming Load, and Exchange show that FiM-Transformer reduces mean squared error by nearly 2.3%, 5.6%, 2.8%, and 3.1% compared with S-Mamba, WPMixer, iTransformer, and PatchTST, respectively, and achieves a Mean Squared Error of 0.166 on Electricity Consuming Load. The proposed method shows strong accuracy, robustness, and generalization in complex scenarios.

摘要： 针对长时序预测中Transformer 类方法计算开销较大、结构适应性不足的问题，提出融合频域先验与线性状态空间模型的预测框架 FiM-Transformer。方法上，首先基于倒置维度策略将多变量历史序列重构为变量主导的词元表示；随后构建多视图超级词元化机制，并行建模全局时域、频域谱特征与局部演化信息，通过自适应门控机制实现多视图特征的协同融合，并引入基于结构风险最小化的灵活骨干网络与动态趋势分解策略，对趋势项与季节、残差项进行异构建模，根据数据噪声水平与预测视域自适应调节模型容量与分解状态，以提升模型对不同数据复杂度的适应能力。实验结果表明，在电力变压器温度、电力消耗负荷和汇率基准数据集上，FiM-Transformer相比S-Mamba、WPMixer、iTransformer和PatchTST分别实现了约2.3%、5.6%、2.8%和3.1%的均方误差降低，其中在电力消耗数据集上取得0.166的均方误差。所提方法在复杂场景下具有较好的预测性能、鲁棒性与泛化能力。