STACnovGRU: weather forecasting based on spatio-temporal adaptive convolutional GRU

Due to the complex spatio-temporal correlation of meteorological data, weather forecasting is a challenging task. Recently, with plenty of meteorological data available and the successful applications of deep learning technology in many areas, developing data-driven models for this task has achieved great attention. Especially, Convolutional Recurrent Neural Networks (CRNNs) have been shown to be effective in spatio-temporal predictive learning. The convolutional connection with shared weights is fixed for different spatial locations and timestamps, while spatio-temporal transformations of meteorological data are varying in both time and space. To address this problem, we developed a Spatio-Temporal Adaptive Convolution for the Gated Recurrent Unit (GRU) to improve the ability of extracting spatio-temporal features. For convenience, we abbreviate our model as STAConvGRU for weather forecasting. The key motivation behind STAConvGRU is to develop additional convolution layers under the framework of the ordinary RNN to learn simultaneously the sampling positions and weights of convolutional kernels. As a result, the adaptive convolution could select the positions and adjust the weights according to the spatio-temporal information. Comparative experiments are conducted on four types of meteorological datasets, including temperature, relative humidity, wind, and radar echo. The experimental results demonstrate the effectiveness and superiority of our proposed model.