Research And Application Of Deep Learning In Regression Prediction

Artificial neural network(ANN)is a kind of machine learning model,which is structed by simulating the information processing mechanism of human brain neurons.ANN has the advantages of self-learning,adaptive and high fault tolerance.However,with the increase of the number of hidden layers in the artificial neural network,there is vanishing gradient problem in ANN,which leads to the optimization of the ANN with multiple hidden layers difficult.In 2006,Hinton et al.introducted a greedy layer-wise unsupervised pre training strategy,which made the training of deep neural network possible.Thus the era of deep learning began.The structure of multiple hidden layers of deep learning could highly abstract the low level features,and realize the automatic extraction of features.In this paper,the deep learning algorithm for regression prediction was studied and it was applied to the short term prediction of wheat aphid.The main research contents are as follows:(1)The greedy layer-wise unsupervised pre training strategy plays an important role in the pre training of deep learning,but the supervising information provided by the sample label is still irreplaceable.The advantage of unsupervised learning is not obvious when dealing with the regression problem with small amount of data.Stacked denoising autoencoder based on supervised pre-training was proposed by changing the feature extraction model of SDAE to supervised learning,which could make full use of the supervising information provided by the sample label and make up the disadvantage of insufficient data.Through the experiments on the UCI regression data sets,the experimental results show that the improved SDAE has smaller prediction error and stronger generalization ability.(2)The setting of hyper parameters has a great influence on the prediction performance of the deep learning model,but at present there is a lack of clear theoretical guidance for the setting of the hyper parameters.On the basis of grid search algorithm,there are two improved algorithms of grid search,block grid search and block changing grid search,which are improved from two aspects of efficiency and accuracy.In order to verify the effectiveness of the improved algorithms,there are some experiments on the UCI regression data sets.The experimental results on Concrete dataset show that the performance of the block grid search and block variable grid search are better than grid search on time,and the prediction ability of SDAE3 is better than SDAE2.Experiments on the Slice dataset show that the MSE of the SDAE0 is 66.19,and the MSE of the SDAE2 which optimizes its hyper parameters by block grid search is reduced to 2.98.The experimental results of Housing dataset show that the prediction results of DBN2 and DBN3 which optimize its hyper parameters by block grid search and block variable grid search are better than the support vector regression and local support vector regression(LSVR).(3)Wheat aphid is one of the main pests which harms the yield and quality of wheat.It can prevent and reduce the loss of wheat by predicting the occurrence of wheat aphids timely and accurately.In this paper,DBN_LSVR was proposed by combining the deep belief network and local support vector regression,and the DBN_LSVR model was used to predict the occurrence of wheat aphids.DBN is mainly used in the feature learning stage of wheat aphid,and the LSVR is mainly used to construct the regression model.The experimental results show that the prediction error of aphid number of the DBN_LSVR is 649.2,and the prediction accuracy of the occurrence degree is up to 83.33%.By comparing with LSVR and DBN model,the experimental results show that the DBN_LSVR model is better than other comparison models.DBN_LSVR provides a feasible scheme for wheat aphid and other pests forecast.