Time Series Prediction Model Based on Selective State Space

针对时间序列数据特征复杂、模式多样且长期依赖导致模型计算复杂度高等问题，文章提出了一种基于选择性状态空间与时间卷积网络的多尺度时间序列预测模型。该模型通过专家混合网络为不同类型的时间序列数据动态分配专家模块，并结合一维离散小波变换与时间卷积网络捕捉多尺度特征。具体而言，专家模块利用小波变换将时序数据分解为高频和低频分量，分别提取短期波动和长期趋势信息，并通过时间卷积网络进行特征提取。处理后的数据输入编码器，采用基于选择性状态空间的Mamba模型同时建模正向和逆向时序信息，增强时序建模能力。实验结果表明，该模型在交通、经济、天气和电力等领域的多变量时间序列预测任务中均显著优于现有基线模型，验证了其有效性与鲁棒性。To address the challenges of complex time series data characteristics, diverse patterns, and high computational complexity caused by long-term dependencies, this paper proposes a multi-scale time series prediction model based on selective state space and temporal convolutional networks. The model dynamically allocates expert modules to different types of time series data through a mixture of expert networks and captures multi-scale features by integrating one-dimensional discrete wavelet transform with temporal convolutional networks. Specifically, the expert modules decompose time series data into high-frequency and low-frequency components using wavelet transform, extracting short-term fluctuations and long-term trends, respectively, and performing feature extraction through temporal convolutional networks. The processed data is then fed into an encoder, which employs a selective state space-based Mamba model to simultaneously model forward and backward temporal information, enhancing temporal modeling capabilities. Experimental results demonstrate that the model significantly outperforms existing baseline models in multivariate time series prediction tasks across various domains, including transportation, economics, weather, and electricity, validating its effectiveness and robustness.