Temperature Prediction Based on Integrated Deep Learning and Attention Mechanism

Xuan Zhao, Lvwen Huang, Yanming Nie

Abstract: It is greatly significant to predict air temperature accurately for effective warning of extreme weather events, whereas the complex and nonlinear characteristics of meteorological data make this kind of forecast difficult to achieve high accuracy. To deal with this issue, a novel model named CNN-GRU-RPASM (Convolution Neural Networks - Gated Recurrent Unit - Relative Position-based Self-Attention Mechanism) was proposed in this paper. Apart from the traditional counterparts, the CNN-GRU-RPASM model innovatively combines the advantages of CNN and GRU, and introduces a gaussian amplifier model to improve the self-attention mechanism with relative position information. Firstly, CNN was used to extract the characteristics of the meteorological input data. Then, the improved self-attention mechanism was employed to extract key information from the data sequence. And finally, GRU was utilized to encode the relationship information among time-series data. The performance evaluation with the real meteorological data shows that the CNN-GRU-RP ASM model performs better than its traditional counterparts. This new model will be deployed in the agricultural production service system to provide technical supports for extreme weather disaster warning forecasting.