| The strong wind has a significant impact on the safe running of trains,and it is one of the most important natural factors that need to be paid attention in the safe running of trains.By the end of 2018,high-speed rail operation mileage of china has reached 29,000 kilometers,which more than two-thirds high-speed rail operation mileage of the world.There are many railway lines across harsh windy areas,such as Lanxin-Tibet and Qinghai-Tibet.The safe running of trains poses a serious threat,because aerodynamic lift and lateral force caused by strong wind.It is of great significance to carry out wind speed prediction research in key areas along the railway for the safety operation,operation scheduling and ride comfort of trains in the whole region.In this paper,a hybrid prediction model based on deep neural network algorithm for meteorological feature extraction is proposed,which is based on the measured data of wind monitoring points in the wind area around Lanxin Passenger Dedicated Railway,according to the realistic demand of wind speed prediction for the safe running of Lanxin Passenger Dedicated Railway Train.The specific research work is as follows:(1)Data preprocessing.Firstly,the wind speed and multi-dimensional meteorological characteristic data of the original input variables are pre-processed,and the rationality and completeness of the original input variables are detected in view of the data loss,uneven distribution and errors of the original data sets.Then,the detected abnormal data are classified,and filled with non-linear regression and mean values.Three methods are used for the wind speed data with partial uneven distribution.The sub-natural spline interpolation method is used to fit the data.Finally,the correlation of multi-dimensional meteorological data is analyzed by mutual information theory.(2)Research on prediction framework of deep neural network model.In the specific process,the historical wind speed is used as input,and two kinds of deep convolution neural network frameworks are used to train the model.Through the deep structure of the network,the intrinsic relationship between the sequences is learned to realize the prediction of the future wind speed series.Be validated,the wind speed prediction errors of CNN+MLP and CNN+LSTM models are at a small level in most cases,and the prediction effect is better than that of traditional weather prediction.Moreover,CNN+MLP depth convolution neural network model performs better than CNN+LSTM model in predicting data sets,has higher prediction accuracy,has better acceptance to abnormal data,and makes the model generalization ability better.Stronger.(3)Wind speed prediction model based on Meteorological feature extraction.Because meteorological features have a great impact on wind speed and the input meteorological features are more and difficult to extract,a CNN+MLP depth neural network model based on multi-dimensional meteorological features extraction is established.Firstly,multi-dimensional historical meteorological data are collected through multiple channels;The optimal meteorological features are extracted from multi-dimensional meteorological features by principal component analysis(PCA),and the optimal meteorological features and historical measured wind speed series are used as input of the prediction model.Then,short-term wind speed prediction models under different meteorological conditions are established by using CNN+MLP depth neural network.The prediction error of the network prediction model is large at the peak of wind speed fluctuation and the prediction accuracy is volatile.The rough value theory is proposed to modify and compensate the prediction value of the model,so as to further improve the prediction accuracy.Experiments show that the short-term wind speed model with the input of the optimal meteorological characteristics has better prediction accuracy and faster calculation speed,which has a certain engineering reference value. |
PDF Full Text Request