D3AT-LSTM: An Efficient Model for Spatiotemporal Temperature Prediction Based on Attention Mechanisms

Accurate temperature prediction is essential for economic production and human society’s daily life. However, most current methods only focus on time-series temperature modeling and prediction, ignoring the complex interplay of meteorological variables in the spatial domain. In this paper, a novel temperature prediction model (D3AT-LSTM) is proposed by combining the three-dimensional convolutional neural network (3DCNN) and the attention-based gated cyclic network. Firstly, the historical meteorological series of eight surrounding pixels are combined to construct a multi-dimensional feature tensor that integrates variables from the temporal domain as the input data. Convolutional units are used to model and analyze the spatiotemporal patterns of the local sequence in CNN modules by combining them with parallel attention mechanisms. The fully connected layer finally makes the final temperature prediction. This method is subsequently compared with both classical and state-of-art prediction models such as ARIMA (AR), long short-term memory network (LSTM), and Transformer using three indices: the root mean square error (RMSE), the mean absolute error (MAE), and the coefficient of determination (R2). The results indicate that the D3AT-LSTM model can achieve good prediction accuracy compared to AR, LSTMs, and Transformer.