融合时序卷积网络与多头自注意力的时间序列预测模型OACHSSCD

Traditional forecasting models exhibit notable limitations in capturing long-term dependencies and enhancing the accuracy of multi-step predictions.To further address issues such as gradient vanishing during long-sequence training,low computational efficiency,and error accumulation in multi-step forecas-ting,this paper proposes a time series forecasting model that integrates temporal convolutional networks,a multi-head self-attention mechanism,and a sequence-to-sequence architecture.First,temporal convolu-tional networks are employed to extract local temporal features,leveraging their dilated convolution struc-tures and residual connections to mitigate gradient vanishing and support parallel computation.Then,a multi-head self-attention mechanism is introduced to model global dependencies and enhance contextual semantics based on the outputs of the temporal convolutional networks,forming a"local-global"collabo-rative feature extraction mechanism.Finally,a sequence-to-sequence framework equipped with an atten-tion mechanism maps the enhanced feature sequences into future multi-step price sequences.Experimental results based on daily historical datasets of multiple A-share stocks in the Chinese stock market demon-strate that the proposed model significantly outperforms baseline methods across multiple metrics,with notable reductions in both mean squared error and mean absolute error,validating its effectiveness and su-periority in the task of multi-step stock price forecasting.