Study On Prediction Of Multiphase Flow State In Wellbore Based On Deep Learning

Wellbore multiphase flow models are widely used in all phases of production in the petroleum industry,and they occupy an important position in the fields of drilling,fracturing,and oil and gas recovery.From the 1940 s to the present,more than 10 kinds of wellbore multiphase flow models have been proposed.According to their derived principles,they can be classified into empirical,semi-empirical and mechanistic models.However,the empirical and semi-empirical models are only applicable in a certain range due to the limitations of their experimental conditions.They are only applicable in a certain range.The mechanistic model also uses some empirical parameters in order to make the model closed.Its extrapolation capability is also relatively limited.In addition,due to the inadequacy of monitoring means and physical-mathematical understanding,there is no unified wellbore multiphase flow model that can be adapted to all cases so far.In view of the above,finding an agent type with a wide range of applicability becomes a way to solve the above problems.In this paper,a deep learning approach is adopted to establish a deep learning model for multiphase flow in a wellbore,using field measured data sets as training samples,combined with data analysis and optimization algorithms,etc.,and certain results have been achieved.Firstly,by comparing the fully linked neural network and convolutional neural network,we analyze the effects of different activation functions and different learning rates on the training speed and training effect of the fully linked neural network hyperparameters.Secondly,the advantages and disadvantages of different convolutional neural networks are introduced.Finally,by comparing the advantages and disadvantages of different neural networks and combining the multi-phase flow model and data characteristics,the fully connected neural network is selected as the deep learning base model.In Chapter 3,the field-collected data set is firstly processed after missing value interpolation,outlier fitting,duplicate data removal,and data normalization.The processed sample set was analyzed for sample skewness,dispersion,magnitude,and normality,etc.On the sub-basis,the sample set was expanded to meet the neural network training requirements using reject-receive sampling.In Chapter 4,the construction of the model and the selection of the loss function are performed first.Since the neural network model possesses many hyperparameters.The speed of model training and the training results vary greatly for different combinations of hyperparameters,so the hyperparameters of the model are preferred using network search combined with optimization algorithms.An optimal set of hyperparameters is obtained.Finally,the training effects of the model under different hyperparameters of the deep learning model are compared to evaluate the performance of the current hyperparameters.Meanwhile,another set of data is selected to test the generalization ability of the model and verify the reliability and accuracy of the model.It is verified that the accuracy of wellbore pressure distribution prediction can reach more than 90%.