Prediction And Research Of Multi-source Meteorological Observation Data Based On Deep Learning

Meteorological data prediction plays an important supporting role in my country's national production.With the rapid development of science and technology,the level of meteorological observation has been continuously improved,the sources of meteorological data have become more and more diverse,and the amount of meteorological data has grown exponentially.However,traditional meteorological forecasting methods have problems such as low forecasting efficiency and low precision when faced with massive meteorological data.Based on this background,this thesis uses deep learning methods to carry out meteorological forecasting research on ground observation data and radar observation data.In terms of meteorological prediction of ground observation data,for the dataset,we use principal component analysis(PCA)to extract the main feature components,reduce the dimension of the input data of the model,and remove the useless noise data in the dataset.At the same time,the meteorological datasets are mostly time series Data,we use the gated recurrent unit(GRU)in the recurrent neural network which has good performance in time series processing,and finally use the exponentially descending inertia weight and the improved particle swarm algorithm(PSO)of the boundary mutation operator to optimize the GRU neural network.Based on this,A recurrent neural network time series prediction model based on principal component analysis and improved particle swarm optimization optimization of gated recurrent units is established.Through empirical analysis,it is confirmed that the model is feasible and accurate.In terms of meteorological prediction of radar observation data,compared with ground observation data,radar observation data are mostly images,and CNN neural network has great advantages in image feature extraction.At the same time,radar chart forecasting precipitation is also a time series forecasting precipitation.Therefore,this thesis constructs a short-term precipitation prediction based on the ConvGRU model by combining the convolutional neural network and the GRU network.Through experimental comparison,although the model achieves good accuracy,there is an over-fitting problem.In order to improve this problem,and propose a prediction model that uses residual structure to improve ConvGRU(Res-ConvGRU).The experimental results show that the new model effectively solves the problem of overfitting and has better performance than other precipitation prediction models.