Research On Meteorological Prediction Based On Long Short-term Memory Network

In recent years,the explosive growth of meteorological data has made the meteorological forecasting field face arduous challenges.Traditional methods can not cope with such massive meteorological data,and the rapid development of deep learning provides new ideas and new ways for the study of meteorological prediction.Most of the meteorological data is time-series data.Long short-term memory network(LSTM),as a thriving technology,can fully consider the temporal correlation in meteorological data and discover the hidden meteorological laws.This paper mainly analyzes the two problems of temperature prediction and short-term rainfall prediction.On the issue of temperature prediction,the ground observation data collected by 20 meteorological elements needs to excavate the elements highly correlated with the temperature.On the other hand,it is necessary to use the meteorological data of historical moments to predict the temperature in the future.A prediction model based on random forest and long short-term memory network(RF-LSTM)is constructed,and good results are obtained which are superior to other models.When performing short-term rainfall prediction,processing radar image data cannot analyze each frame alone,but also analyze the entire time series in which the frames are connected,and take the spatial characteristics of the radar image into account.Therefore,combing convolutional neural network with LSTM,a predictive model based on convolutional long short-term memory network(ConvLSTM)is constructed,which is compared with the experimental results of two-dimensional convolution(Conv2D)and three-dimensional convolution(Conv3D)models.It was found that the predictive effect of ConvLSTM has a clear advantage.In order to further improve the accuracy of short-term rainfall,a predictive model based on Conv3 D and bidirectional long short-term memory(Conv3D-BiLSTM)is proposed by improving the network structure.Experiments show that the Conv3D-BiLSTM model is superior in both network convergence and final prediction accuracy.