Prediction For Shale Gas Production Based On Mass Data Technology

The exploitation of shale gas is an important policy of the national energy development strategy and an important way to ensure the energy security of China.However,it is not advisable to blindly develop shale gas wells.For the high benefit of shale gas development,production prediction of shale gas wells has become a hot region in the oil and gas field.But the current research on production forecast often needs to establish a mathematical model under ideal conditions first,which makes its forecasting accuracy in practical applications not high.Meanwhile,with the development of shale gas wells,a large amount of geological exploration data,construction data and production data have been accumulated.Therefore,this thesis constructs a big data system based on distributed storage,and constructs a shale gas production prediction model based on deep neural network,trying to solve the problem of production prediction.The contribution and summary of this thesis are summarized as follows:1.This thesis constructs a big data management and analysis system based on distributed storage,which realizes the automatic analysis function of shale gas original data files in different formats,and provides the infrastructure for the shale gas data sets and production prediction models.2.In this thesis,a deep neural network production prediction model based on residual network is proposed.The residual network is used to extract geological characteristics,and combined with construction characteristics and production characteristics to predict the production of shale gas.Compared with the traditional machine learning models such as support vector machine,logistic regression and decision tree,the prediction ability of production prediction model is greatly improved.3.Based on the production prediction model based on residual network,this thesis introduces the attention mechanism based on channel,calculates the importance of each geological feature channel in combination with construction characteristics and production arrangement characteristics,and then improves the geological features useful for shale gas production prediction and suppresses the geological features of little use according to this importance.This further improves the accuracy and stability of shale gas production prediction.4.Then,based on the label based online knowledge distillation framework,this thesis transforms the training of yield prediction model.That is,by constructing a multi branch network,each branch network can learn the label knowledge of the teacher network and improve the training effect of the model.Experiments show that the online knowledge distillation framework can improve the prediction performance of the model without changing the network structure of model inference.5.Finally,according to the characteristics of the label of the shale gas dataset,Wasserstein distance is used to enhance the measurement of the distillation loss in the online distillation framework so as to improve the training effect of online knowledge distillation.