Application Research Of Recurrent Neural Network In Sand-Dust Storm Forecast In Inner Mongolia

Sand-dust storm includes sand storm and dust storm,often occurs with desertification.It refers to a kind of weather phenomenon,which is sudden and harmful,that the unstable strong airflow brings up the dust on the ground,which leads to the decrease of visibility.Sand-dust storm occur more frequently in northwest and north China.Inner Mongolia has a long and narrow terrain,spanning three regions,includes northeast China,north China and northwest China,and the central and western regions of Inner Mongolia are the areas with frequent sand-dust storm.Sand-dust storm control mainly through the restoration of vegetation,to protect the environment;In terms of prevention,methods such as sandstorm warning and forecasting are needed.In recent years,the analysis and modeling of meteorological data by computer has gradually become a research hotspot.Recurrent Neural Network(RNN)is a kind of deep learning model,also knows as Time Recursive Neural Network.Due to meteorological data is a typical time series with multi-dimensional attributes,uncertainty,spatiotemporal characteristics and periodicity,it is difficult to obtain good accuracy by processing and analyzing meteorological data with traditional data analysis methods.Recurrent Neural Network can model the time series well,and it is suitable for the analysis and processing of large quantities of meteorological data.Therefore,this project adopts to use Recurrent Neural Network model to analyze and modeling the meteorological data,forecast the weather phenomenon of sand-dust storm in Inner Mongolia,and provide a certain degree theoretical support for improving sand-dust storm warning level.This project mainly completes the following work:First,the collected sand-dust storm data and the corresponding daily meteorological data were integrated and pretreated.It mainly includes: data integration,data dimension reduction,handling eigenvalue values and handling measurement missing values,etc.It was arranged into time series according to the time of data acquisition,and the imbalance of data was treated by using Synthetic Minority Oversampling Technique(SMOTE).Second,the advantages and disadvantages of Recurrent Neural Network are analyzed,and the structure and parameters of Long Short-Term Memory(LSTM)model and Gated Recurrent Unit(GRU)model are determined.Using the above two neural networks and the pre-processed meteorological data sequence,the sandstorm forecast model is established,and the forecast results of the two models are analyzed and compared.The experimental results show that the LSTM neural network has better sand-dust storm forecast performance and is more suitable for this project.Finally,aiming at the shortcomings of the above LSTM model and its forecast,the model of Convolutional Neural Network(CNN)based on time series data is established,and the Stacking ensemble strategy is adopted to ensemble the two kinds of neural networks.Two RNN-CNN ensemble sand-dust storm forecast models were established by using the time series memory capability of Recurrent Neural Network and the local feature extraction capability of Convolutional Neural Network,respectively,with fully connected network and support vector machine as the meta-classifier.The experimental results show that the ensemble model with fully connected network as the meta-classifier has better forecast performance.Compared with the single RNN model and the single CNN model,this ensemble model has improved to different degrees in a variety of forecast evaluation indexes,including accuracy,precision,recall and f1-score.