SA-JSTN: Self-Attention Joint Spatiotemporal Network for Temperature Forecasting

The rapid development of remote sensing technology has brought abundant data support for deep learning based temperature forecasting research. However, recently proposed methods usually focus on the temporal relationship among observation information, whereas ignore spatial positions different regions. Motivated by that adjacent regions present similar trends, in this paper we consider as a spatiotemporal sequence prediction problem, and propose new model forecasting, Self-Attention Joint Spatiotemporal Network (SA-JSTN), which simultaneously captures interdependency information. kernel component SA-JSTN is newly developed Memory (STM) unit, describes models via unified memory cell. STM constructed units convolutional LSTM (ConvLSTM). Instead using simple convolutions information extraction, improve ConvLSTM self-attention module, significantly enhanced global representation ability our network. Compared with other methods, able to integrate correlation into series thus better performance especially short-term prediction. We have conducted comparison experiments two typical sets validate effectiveness method.