Research On Meteorological Element Prediction Based On Cross Attention And Recurrent Neural Networ

With the development of meteorological observation methods,meteorological observation data has become increasingly abundant,providing richer information for meteorological forecasting tasks.However,traditional data assimilation numerical forecasting requires extensive computational resources and experienced meteorological forecasters,which poses bottlenecks and challenges for utilizing multiple meteorological data sources in meteorological forecasting tasks.The development of artificial intelligence technology provides a new approach to solving the above problems.This paper proposes a multimodal data fusion based on the cross-attention model for short-term precipitation prediction and a Wavelet Self-Attention Net model based on 2D wavelet decomposition and attention mechanism for daily sea surface temperature prediction.The specific contents are as follows:The first part of this paper proposes a short-term precipitation forecasting method based on a cross-attention mechanism.Firstly,meteorological station data and radar echo data in Guangdong and Guangxi provinces are processed to establish a multimodal short-term precipitation forecasting dataset.Secondly,a recurrent neural network is used to design the station data feature extraction module and radar echo data feature extraction module.Then,the cross-attention mechanism is utilized to align and exchange the station and radar features.Finally,a multi-layer perceptron is used to output short-term precipitation values,thereby improving the accuracy of precipitation forecasting.Through experimental comparative analysis,this method achieved good results in historical precipitation forecasting in Guangdong and Guangxi provinces and has reference value.The second part of this paper proposes a novel end-to-end dual-branch sea surface temperature prediction model.The left branch adopts a wavelet attention module to extract wavelet features in the spatiotemporal sequence using 2D discrete wavelet decomposition and attention mechanism to enrich the structural information and achieve multi-frequency analysis.The right branch adopts a Convolutional Long Short-Term Memory to learn complementary features of wavelet features in the spatiotemporal sequence and compensate for the feature information the left branch fails to learn.Through experimental comparative analysis,the proposed model surpasses or reaches the level of current advanced models in multiple indicators.Furthermore,the model’s effectiveness in terms of convergence speed and prediction image quality is analyzed through ablation experiments.