基于多尺度并行架构卷积的两阶段多元时间序列预测模型OA

Modeling multi-periodic and seasonal patterns in time series data is complex.Meanwhile,difficulties in fusing multi-scale structures and significant differences in variable interaction intensities across domain,led to model forecasting accuracy is limited and cross-scenario adaptability is insufficient.To address these issues,this paper proposes a two-stage multivariate time series forecasting model based on a multi-scale parallel architecture convolution.The model adopts a progressive two-stage learning strategy combining channel independence and channel mixing,sequentially modeling the dependencies within individual channels and across channels to enhance the expressive capability for data from diverse scenarios.Its core consists of multiple parallel branches incorporating deep convolutions and dilated convolutions,which extract features of multi-scale periodicities by configuring differentiated receptive fields.Additionally,a convolutional attention module is integrated to strengthen feature representation and improve model accuracy.Finally,a dual prediction head structure is employed to fit both linear and non-linear relationships in the data,and the results from both heads are fused to generate reliable predictions.Comparative experiments with eight advanced methods on seven real-world data-sets demonstrate that the proposed model achieves an average reduction in MSE by 4.4％,12.1％,2.5％,7.9％,9.3％,4.8％,1.3％and 13.6％respectively,effectively improving prediction accuracy.