A Heterogeneous Spatiotemporal Attention Fusion Prediction Network for Precipitation Nowcasting
Dan Niu, Hongshu Che, Chunlei Shi, Zengliang Zang, Hongbin Wang, Xunlai Chen, Qunbo Huang
Abstract
Precipitation nowcasting underlying various public services from rainstorm warning to flight safety, is quite important and remains challenging due to the fast change in convective weather. Although some deep learning models have been proposed to make prediction automatically, most of them just deal with a single radar echo data source, making them hard to adapt to heterogeneous and diverse data in practice. In this work, a heterogeneous spatiotemporal attention fusion prediction network (HST-AFP) is proposed for radar echo extrapolatio(deterministic output) and further precipitation nowcasting, which deals with mining and fusing knowledge from multiple heterogeneous spatiotemporal (ST) data sources, including history radar echo observations and numerical weather prediction (NWP) data. With the help of the proposed attention-based ST diffusion module (ASTD), the multi-encoder is designed to extract information from both dense ST tensor and sparse ST tensor. On the other hand, the fusion-decoder achieves very deep trainable residual fusion prediction by integrating scale-wise attention fusion module (SWAF) and deep residual spatial and temporal attention mechanism (DRSTA). It can adaptively blend multi-source ST features and rescale the multi-scale temporal-wise and spatial-wise features for better prediction. Experiments in a real-world dataset of South China show that compared with the ingenious RNN-based methods and newly proposed Unet-based methods, our HST-AFP network can handle complex input with heterogeneity in both space and time domains, and performs better on the precipitation nowcasting metrics, as well as requires remarkable shorter forecast time.