【摘要】 时间序列预测在金融、医疗、交通和气象等领域发挥着重要作用。在长时间序列预测中，迫切需要提高预测的精度，解决内存不足等问题。近年来，Transformer模型在自然语言处理领域得以成功应用的同时，在预测研究领域也引起了学者们的广泛关注，Transformer变体Informer模型的研究在时间序列预测中取得了较大进展。本研究以Informer框架为基础，与加性注意力机制相结合，提出了A~2former模型。利用A~2former模型在ETT,WTH,ECL和PM2.5数据集上进行了长时间序列预测的实验，实验结果表明所提模型在长时间序列预测中表现出比基线方法(如Informer模型和LSTMa模型)更好的性能。A~2former模型不仅将计算时间复杂度降低到线性，而且可以实现更有效的序列建模。本研究的工作为时间序列预测提供了有益参考。更多还原

【Abstract】 Time series forecasting plays an important role in the fields of finance, medicine, transportation and meteorology. In long sequence time-series forecasting(LSTF), it is urgent to improve the forecast accuracy and solve the problems of insufficient memory. In recent years, the successful application of Transformer in natural language processing has also attracted a lot of attention forecasting studies. Informer model, a variant of Transformer, has made great progress in time series forecasting. In this paper, we proposed a A~2former model, which is based on Informer and additive attention mechanism. The A~2former model was experimented on ETT, WTH, ECL and PM2.5 datasets for LSTF. Experimental results show that A~2former exhibits better performance than existing baseline methods(e.g., LSTMa and Informer) in LSTF. A~2former not only reduces time computational complexity to linearity and improves forecast accuracy, but also enables more efficient sequence modeling, our work provides a valuable input for time series forecasting.更多还原