| In recent years,the application of machine learning in the dynamic analysis of oil and gas reservoir has become more and more extensive,which can predict the production of oil and gas reservoirs under different conditions more quickly and accurately,and improve the economic benefits of oil and gas exploitation ultimately.In this paper,in terms of the production prediction problems in different situations in the production process of oil and gas reservoirs,the research of different machine learning models is carried out,and the main tasks completed are as follows:(1)A machine learning method based on static geological and engineering factors to predict the EUR of gas wells has been established.Firstly,single-factor and multi-factor analysis of factors affecting the EUR of gas wells are carried out,and the relationship between the factors is clarified.Then,the grey relational analysis and regression analysis are established with the first year output as the intermediate quantity.The Mean Absolute Percentage Error(MAPE)of new model is 9.11%,which is better than prediction performance of multiple linear regression model.Finally,8parameters including pressure coefficient,the thickness of a small layer,main fracturing fluid volume,brittle mineral content,drilling length,fracturing length,TOC and gas content,are selected as the main factors affecting the EUR of gas wells from the 12 parameters.(2)The Linear Dynamical System(LDS)production prediction model for water drive oil and gas reservoirs is established.And,the model parameters are estimated by the EM method.The model is applied to the examples of block F in China and four blocks of oil and gas fields in the Nevada area of the United States.The MAPE of the prediction results of four blocks in the United States are less than 9% and the MAPE of the prediction result of liquid production in block F is equal to 13.49%,which shows the feasibility of the model.The correlation between input parameters and prediction accuracy is discussed.When the model input data satisfies the approximate Gaussian distribution(normal distribution)and the input and output data have similar trends,the accuracy of the prediction results is higher.(3)A BP-LSTM production forecast model is established.The model uses a total of 14 static parameters of geology and engineering to make non-linear predictions of new well dynamic monthly production data.Research shows that the MAPE of the prediction results is about 20%,which is smaller than the MAPE of the prediction results using the BP model.And the comparison results with the accuracy of the linear prediction model in Chapter 2 show that the BP-LSTM model has a more accurate fitting ability to nonlinear data. |
