Analysis Of Airport Atmospheric Visibility And Evolution Law Based On Deep Learning

Airport visibility refers to the distance at which the human eye can clearly see the runway and surrounding ground objects.It is an important meteorological parameter in aviation transportation and is usually affected by atmospheric factors such as rain,fog,snow,and sandstorms.It directly affects the safety of aircraft takeoff and landing.Deep learning,with its efficient feature learning capability,strong generalization ability,and high scalability,enables prediction tasks by learning feature representations and classification decisions from data.Therefore,this study focuses on the prediction of airport visibility and its evolution based on deep learning.The main research contents are as follows:(1)Selecting key meteorological factors for airport visibility prediction.Meteorological observation data from Zhengzhou Airport station of the China Meteorological Administration were used.The data were analyzed,and missing values were filled using linear interpolation.Principal component analysis(PCA)was employed to determine the attributes of the input data set,including PM2.5 concentration,PM10 concentration,CO concentration,temperature,humidity,and visibility indicators.(2)To address the issues of insufficient feature extraction and potential overfitting in traditional prediction models,this study proposes the LSTM-AED(LSTM-AutoencoderDropout)model for airport visibility prediction.Firstly,the LSTM model is used as the foundation,and improvements are made through data preprocessing and parameter optimization.Next,to tackle the problem of inadequate data extraction by the LSTM model leading to increased prediction deviation in long forecasting steps,an LSTM-AE(LSTMAutoencoder)structure is designed.The LSTM-AE model effectively extracts features from different data characteristics,as verified by experimental results.Furthermore,in order to enhance the prediction accuracy of airport visibility,the issue of overfitting is addressed by introducing the random dropout mechanism in the LSTM-AE model.This results in the LSTM-AED prediction model,reducing the reliance on specific neurons and effectively avoiding overfitting.Experimental results demonstrate that the LSTM-AED model achieves good prediction performance,although some time delay issues are observed in the model predictions.(3)In response to the time delay issue in predicting airport visibility using the LSTM-AED model,this study proposes the LSTM-AED-BA(LSTM-Autoencoder-DropoutBahdanau Attention)model for the evolution prediction of airport visibility,based on the Bahdanau Attention mechanism.This model adaptively adjusts the importance of each position by learning attention weights,thereby increasing the model’s focus on key information,enhancing flexibility,accuracy,and reducing prediction time delay.Once the key parameters of the model are determined,to further enhance the model’s generalization ability,the Adam optimizer is combined with an adaptive learning rate decay strategy,resulting in the design of the R-Adam optimizer.This dynamic adjustment of model parameters allows better adaptation to different training scenarios.Experimental results demonstrate that the LSTM-AED-BA model effectively mitigates the time delay issue in predictions and achieves higher prediction accuracy compared to the LSTM-AED model.This validates the generalization and superiority of the LSTM-AED-BA model in airport visibility prediction.