Abstract: Abstract: Time series data come from a wide range of social fields, from meteorology to finance to medicine. Accurate long-term prediction is a key issue in time series data analysis, processing, and research. Based on the correlation of different scales in time series data, a multi-scale information fusion time series long-term forecasting model based on neural network (ScaleNN) was proposed, which aims to better handle the multi-scale feature in time series data to achieve more accurate long-term forecasts. The architecture combines full-connection neural networks and convolutional neural networks to efficiently extract global and local information and aggregate the two for prediction. In addition, by introducing a compression mechanism in the global information representation module, the ScaleNN model can accept longer sequence inputs with a lighter structure, which increases the perceptual range of the model and improves the model's performance. Experiments demonstrate that ScaleNN outperforms the current best-in-class model on multiple real-world datasets while improving runtime by 35% and requiring only 19% of its number of parameters.

Key words: time series, big data, data mining, deep learning, series forecasting

摘要： 时间序列数据广泛来源于社会各个领域中，从气象学到金融学再到医学，准确的长期预测是时间序列数据分析、处理与研究中的一个关键问题。针对时间序列数据中存在的不同尺度相关性的挖掘与利用，本文提出了一种基于神经网络的多尺度信息融合时间序列长期预测模型（ScaleNN），旨在更好地处理时间序列数据中的多尺度问题，以实现更准确的长期预测。该模型架构结合了全链接和卷积神经网络，能够有效地提取全局信息与局部信息，并将两者聚合后进行预测。此外，通过在全局信息表征模块中引入压缩机制，ScaleNN模型可以以更轻量化的结构接受更长的序列输入，增大了模型的感知范围并提高了模型效能。实验证明，ScaleNN在多个真实世界数据集上的性能优于当前领域最佳模型PatchTST，同时运行时间提升35%并仅需其19%的参数量。

关键词: 时间序列, 大数据, 数据挖掘, 深度学习, 序列预测