Imputing Missing Temperature Data of Meteorological Stations Based on Global Spatiotemporal Attention Neural Network

Imputing missing meteorological site temperature data is necessary and valuable for researchers to analyze climate change and predict related natural disasters. Prior research often used interpolation-based methods, which basically ignored the temporal correlation existing in the site itself. Recently, researchers have attempted to leverage deep learning techniques. However, these models cannot fully utilize the spatiotemporal correlation in meteorological stations data. Therefore, this paper proposes a global spatiotemporal attention neural network (GSTA-Net), which consists of two sub networks, including the global spatial attention network and the global temporal attention network, respectively. The global spatial attention network primarily addresses the global spatial correla-tions among meteorological stations. The global temporal attention network pre-dominantly captures the global temporal correlations inherent in meteorological stations. To further fully exploit and utilize spatiotemporal information from me-teorological station data, adaptive weighting is applied to the outputs of the two sub-networks, thereby enhancing the imputation performance. Additionally, a progressive gated loss function has been designed to guide and accelerate GSTA-Net's convergence. Finally, GSTA-Net has been validated through a large num-ber of experiments on public dataset TND and QND with missing rates of 25%, 50%, and 75%, respectively. The experimental results indicate that GSTA-Net outperforms the latest models, including Linear, NLinear, DLinear, PatchTST, and STA-Net, across both the mean absolute error (MAE) and the root mean square error (RMSE) metrics.