Research On Temperature Prediction Based On Graph Convolutional Neural Networ

Weather prediction is an important modeling task,which has a significant impact on agriculture,industrial production,transportation,military fields,and people’s daily lives.In recent years,meteorological observation technology has been continuously developed,meteorological data has become increasingly rich,and the rise of deep learning has provided new solutions for weather prediction research.Temperature is one of the most important elements in meteorological data,Meteorological data exhibits obvious spatiotemporal characteristics and complex nonlinear relationships.Traditional neural network models such as Long Short Term Memory can take into account the temporal characteristics of meteorological data,but ignore the spatial relationships in meteorological data,compared to traditional classical sequence prediction models,the graph convolutional neural network temperature prediction model proposed by combining gated cyclic units has higher accuracy in temperature prediction.On the construction of meteorological map,in order to reduce the complexity of model input and accurately capture the connection relationship between nodes,this paper proposes a graph convolution recurrent neural network temperature prediction model based on L1 regularization,which uses L1 regularization method to reduce model parameters.During training,the model can learn the sparse adjacency matrix of the graph,effectively solving the overfitting problem caused by the excessive number of nodes connected in the graph,it improves the efficiency of model operation and has high accuracy.In determining the weights of connections between nodes in the graph,this paper proposes a graph convolutional recurrent neural network temperature prediction model based on graph attention mechanism to study the spatial correlation differences of meteorological maps and distinguish the importance of different meteorological stations to the entire meteorological network.Different learning weights are assigned to different adjacent nodes through graph attention mechanism,and the correlation between meteorological nodes is adaptively determined in the spatial dimension,through comparative experiments,it has been proven that the model incorporating graph attention mechanism has higher robustness and accuracy in temperature prediction.At the same time,this paper also discussed the impact of the value of Chebyshev K on the experimental results.After experimental verification,too many or too few nodes can cause a decrease in model accuracy.Therefore,selecting an appropriate K value is particularly crucial.