首页|Spatiotemporal UNet for multi-field spatiotemporal series generation
Spatiotemporal UNet for multi-field spatiotemporal series generation
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
Taylor & Francis
ABSTRACT Building multi-field spatiotemporal virtual environments is of great importance for industrial applications. However, due to the calculation complexity and the lack of consideration for spatiotemporal correlation, existing methods cannot meet the real-time and accuracy requirements. In this paper, we propose a novel Spatiotemporal UNet based on the three-dimensional convolutional neural network for the generation of multi-field spatiotemporal series. The MultiScale Attention Head is proposed to learn the multiscale spatiotemporal information. The Time InfoFusion module is proposed to mine the temporal correlation from spatiotemporal series. Moreover, considering the long-term periodicity and short-term stochasticity of spatiotemporal series, we propose the absolute time encoding strategy and spatiotemporal moving averages mean square error to optimize network learning. Experiments on three different regions of different physical fields show that the proposed method can generate virtual spatiotemporal environments in different tasks, with the RMSE decreasing 0.96–49.59% in 3-day lead time task, 0.16–49.97% in 5-day lead time task, and the MAE decreased up to 54.33% in 3-day lead time task and 55.30% in 5-day lead time task. The inference speed of Spatiotemporal UNet is 4.66 times of ConvLSTM, indicating its accuracy and real-time performance.
NETL
NSTL
Taylor & Francis
