深圳大学学报（理工版）2025，Vol.42Issue(3)：334-341,8.DOI:10.3724/SP.J.1249.2025.03334
基于双模型并联的复杂时序预测方法
Complex time series forecasting method based on dual-model parallelism
摘要
Abstract
Traditional forecasting models typically focus on capturing trends and patterns in time series data while often neglecting interactions among variables,which limits their effectiveness in complex time series scenarios.To address this,a dual-model parallel approach named Dualformer is proposed,which integrates the inverted transformer(iTransformer)and patch time series transformer(PatchTST)models in a parallel architecture.In this framework,the conventional feedforward neural network components are replaced with activation functions,and a multilayer perceptron is introduced to further process the outputs of the two parallel sub-models.The Dualformer model employs attention mechanisms to simultaneously capture both the temporal and variable dimensions in the complex time series,effectively capturing time-dependent patterns and multivariate interactions.Experimental results demonstrate that the Dualformer significantly outperforms the benchmark models,including iTransformer,PatchTST,and decomposition linear(DLinear)in complex time series forecasting tasks.These results highlight Dualformer's superior accuracy and its strong potential for real-world applications in forecasting complex time series data.关键词
人工智能/深度学习/复杂时序预测/注意力机制/多层感知机/Dualformer模型Key words
artificial intelligence/deep learning/complex time series forecasting/attention mechanism/multilayer perceptron/Dualformer model分类
信息技术与安全科学引用本文复制引用
郑洪英,夏林中,刘星..基于双模型并联的复杂时序预测方法[J].深圳大学学报（理工版）,2025,42(3):334-341,8.基金项目
National Natural Science Foundation of China(62303327) （62303327）
Natural Science Foundation of Guangdong Province(KJ2023B001) 国家自然科学基金资助项目(62303327) （KJ2023B001）
广东省自然科学基金资助项目(KJ2023B001) （KJ2023B001）
