Research And Application Of Convolutional Neural Network Modeling

With its advantages of feature extraction and nonlinear data modeling,deep learning has gained more and more attention from different researchers.As a typical structure in deep learning,convolutional neural network has been widely used in the fields of chemical engineering,biology and speech recognition.There are several characteristics such as local connections and shared weights etc.These features can reduce the complexity of the network and the number of training parameters,and they also can make the model having strong fault tolerance.In recent years,the proportion of new energy in China's power industry is increasing,and the power system with a high proportion of new energy is gradually constructed.Therefore,the accurate prediction of wind speed is of great significance to improve the stability of the power system under wind power combination.In this paper,wind speed forecasting is taken as the application object,and wind speed prediction model is constructed based on these superior characteristics of CNN.In the CNN,the convolution layers and the pooling layers have the function of extracting the inherent features hidden in the original data,and the full connection layers have the function of integrating different local information extracted.Moreover,the CNN can be trained by supervised learning,which makes it easier to optimize the parameters.Due to the strong fluctuation and intermittency of wind speed,the filtering and extraction features of the pooling layers are too rough.Therefore,the pooling layer is not used in the CNN model established in this paper which is suitable for wind speed forecasting.Due to the existence of unactivated dead neurons in the convolution layer and the large amount of simulated wind speed data,the Dropout method is applied to each layer of the model to improve the generalization performance of the model.Thus,a wind speed forecasting model based on the CNN with Dropout is established,and the model is trained by the small-batch gradient descent algorithm.To solve the problem of network input data selection,the feature selection method based on chi-square test is adopted.The data of a real wind farm in Inner Mongolia are used for simulation verification.The results prove the validity of the above models,and achieve higher prediction accuracy.With the Dropout method,the dropout rate for each layer is a fixed value set empirically,which may increase the generalization gap.Therefore,the distribution of the dropout rate of each layer is derived by using the Lyapunov stability method while ensuring the convergence of the network.Based on this,an adaptive Dropout method is proposed and applied to the wind speed forecasting model of CNN.And the model does not increase the computational complexity of.Based on a variety of error evaluation indexes,the wind speed forecasting model of the CNN based on the adaptive Dropout method is verified by simulation examples.The results show that the model can not only maintain the prediction accuracy,but also accelerate the convergence speed.