Research On Permeability Prediction Method Of Low-Permeability Reservoir Based On LSTM

Translate the following into English:Reservoir physical parameter evaluation is an important goal of logging data processing and interpretation,among which permeability is used to quantitatively describe the permeability of reservoir rocks,which has important guiding significance for the exploration,development and production of oil and gas fields.Due to the heterogeneity of the reservoir and the complexity of geological conditions,it is extremely difficult to predict the permeability.Low-permeability reservoirs have gradually become the focus of oil and gas exploration and development.The mapping relationship between logging data and reservoir parameters is more complex,and traditional models and methods are difficult to achieve application results.Emerging machine learning methods have powerful nonlinear mapping capabilities and show strong advantages in solving complex nonlinear relationships between data.In view of the characteristics of time series in the depth of the nonlinear regression and well logging data of the permeability prediction task in this paper,the XGBoost network and the general BP neural network in the traditional machine learning method are selected as the basic network,and the LSTM neural network is used for this task.Test and compare the model performance and prediction effect.At the same time,combined with domain knowledge,the LSTM neural network is improved,and the pseudo residual structure is introduced to strengthen the guidance of domain knowledge to the network output,and its model performance and test results are analyzed.The source of the data in this paper is the Chang 6 reservoir section in the Huangling area of the Ordos Basin.Statistical analysis of the data shows that the rock type of the target layer is mainly feldspar sandstone,and the pore types are mainly dissolution pores and residual intergranular pores,with an average porosity of 8.4%.The average permeability is 2.39m D,which belongs to low porosity and low permeability reservoir.On the basis of the quality inspection of the well logging curve,the linear correlation method is used to correct the depth of the well logging curve and the core is returned,and the random forest method is used to identify and deal with the outliers,and the XGBoost built-in module is used to determine the importance of the input features.Based on the analysis and combination of reality,eight logging curves are selected as the input of the network model.According to the characteristics of the data and the task,MSE was selected as the loss function,Re LU and tanh were chosen as activation functions,and Dropout was used for regularization.Optuna and Adam were used to optimize the hyperparameters of the network.Based on the best network parameters,the network was trained and evaluated using five performance metrics,including MAE,MSE,RMSE,R~2,and conformity rate.The results showed that the LSTM neural network model and its improved model,the LSTM-PI neural network model,had the smallest MAE,while the XGBoost network model had the largest MAE.The LSTM-PI neural network model had the smallest MSE and RMSE and the largest R~2,while the XGBoost network model had the largest MSE and RMSE and the smallest R~2.As for the conformity rate,the conformity rates of the four network models(XGBoost,BP neural network,LSTM neural network,and LSTM-PI neural network)were all higher than that of the pore-throat linear regression model,with the LSTM neural network model and LSTM-PI neural network model having the highest conformity rate,and the XGBoost network model having the lowest conformity rate.Based on the comprehensive evaluation of all metrics,the LSTM-PI neural network model had the best prediction performance for permeability prediction in the target layer of the study area,followed by the LSTM neural network model,while the BP neural network and XGBoost network had relatively good and average prediction performance,respectively.The trained models were applied to the test wells in the study area,and the results showed that the LSTM-PI neural network model had the smallest prediction error,which further confirmed the guidance significance of domain knowledge for purely data-driven network models.