摘要
时空序列预测旨在基于历史观测的时空序列数据预测未来一段时间内的状况,时空序列具有复杂的时空相关性,常见方法随着预测步长的增加会丢失长程依赖,导致最后几帧的预测精度大大降低。论文提出一个融合三维卷积神经网络和循环卷积神经网络的混合模型进行时空序列预测,三维卷积神经网络主要捕获长程依赖,提取固定长度历史信息的全局时空特征;循环卷积神经网络用于捕获短程依赖,提取帧间的局部时空特征,通过设计门控单元来融合两种模态的信息,除此之外,采用多尺度输入策略来提高预测图像的清晰度。实验表明,混合模型的预测结果优于常见的预测模型,其中引入的三维卷积模块极大地提高多步预测的精度,降低了预测误差。
Spatio temporal series prediction aims to predict the situation in the future for a period of time based on the spatio-temporal series data of historical observations.Spatio temporal series has complex spatio-temporal correlation,and common methods will lose the long-range dependence with the increase of prediction step size,leading to the greatly reduced prediction accuracy of the last few frames.In this paper,a hybrid model integrating 3D convolutional neural network and cyclic convolutional neural network is proposed to predict time-space series.3D convolutional neural network mainly captures long-range dependence and extracts global time-space characteristics of fixed length historical information.Cyclic convolution neural network is used to capture short-range dependence,extract local spatio-temporal features between frames,and fuse the information of two modes by designing a gating unit.In addition,multi-scale input strategy is used to improve the clarity of the predicted image.The experiment shows that the prediction result of the hybrid model is better than that of common prediction models.The introduction of 3D convolution module greatly improves the accuracy of multi-step prediction and reduces the prediction error.
作者
唐海波
陆振宇
杨强
TANG Haibo;LU Zhenyu;YANG Qiang(School of Artificial Intelligence,Nanjing University of Information Science and Technology,Nanjing 210000)
出处
《计算机与数字工程》
2025年第9期2450-2454,共5页
Computer & Digital Engineering
基金
国家自然科学基金联合重点项目(编号:U20B2061)资助。
