Research On Prediction Method Of Reservoir Physical Property Parameters Based On Ensemble Learning

Reservoir physical property parameters are the basis of reservoir evaluation and are important parameters to describe reservoir characteristics and fluid patterns.The traditional prediction model is based on well logging modeling to fit the porosity and permeability parameters.The main methods include regression analysis and empirical formulas,but most of these methods are linearity.In the actual logging process,the reservoir is complex and heterogeneous,and the logging data may be affected by noise and outliers,which will affect the prediction effect of the model to a certain extent.The traditional prediction model is based on physical model and statistical method,which is easy to overfit the training data and further affect the prediction effect.Therefore,traditional regression analysis and other methods are not enough to meet the demand,and an efficient and accurate forecasting method is urgently needed.The performance advantage of machine learning in nonlinear regression tasks has attracted much attention from researchers at home and abroad,and a lot of in-depth research has been carried out in recent years.By building a model and integrating a large number of input features,researchers attempt to uncover nonlinear relationships between variable features.Therefore,the prediction algorithm of reservoir physical property parameters has gradually changed from the traditional linear regression model to the modeling method on the basis of machine learning.In the field of machine learning,the accuracy and stability of the algorithm can be significantly improved by using ensemble learning method.At the same time,the integration of prediction results of multiple models can effectively reduce model variance and improve model generalization ability,so as to avoid overfitting problemsAccording to the public sample data set,a model that could be used to predict reservoir parameters is established by using precision weighting and sample extension methods to improve the traditional stacking integrated learning in this paper.In view of the large amount of data and the problem of redundant features,the variance method,Pearson correlation coefficient method and model-based feature selection method is used in this study to conduct correlation analysis of sample features,so as to find out the logging curves closely related to porosity and permeability.Firstly,variance method is used to calculate the divergence degree of data,and threshold value is set.After removing characteristic values less than threshold value,thermal map is further used to display Pearson correlation coefficient between reservoir physical property parameters and each logging curve,and the relationship between them and logging curves is separately analyzed to screen out the characteristic data with strong correlation.In order to further improve the model performance,then the feature selection method based on the model is used to select the logging data related to porosity and permeability.Secondly,the data are preprocessed to make the logging data in the same dimensional interval.Founded upon the problems of high modeling cost and multiple factors affecting accuracy of traditional physical parameter prediction model,a fusion algorithm of TCN-LSTM is proposed.The Temporal Convolutional Network(TCN)algorithm and the Long Short Term Memory(LSTM)algorithm are compared respectively to prove the effectiveness of the fusion algorithm.Then,the well logging sample data are substituted into five models for training,namely XGBoost,SVR(Support Vector Regression,SVR),TCN-LSTM,BP(Back Propagation,BP)neural network and ARIMA,and the parameter values are determined by particle swarm optimization algorithm.Then,the above models are used to predict the porosity and permeability of the open sample data set,and the fitting graphs of predicted values and interpreted values as well as the RMSE and MAPE values of each model were analyzed.Grounded in the prediction of physical parameters of a well,comparing the prediction effect of the above model with the traditional stacking integrated model and the improved stacking integrated model in this paper,it is found that no matter from the fitting graph of predicted and explained values or two RMSE and MAPE values of the model,the improved stacking integrated models are the most accurate and effective in predicting the stacking.Finally,the improved stacking integrated model is applied to actual reservoir physical parameters.Experiments are conducted on the actual porosity and permeability parameters of a low permeability well A and well B,and compared with the machine learning algorithms mentioned in this paper and the traditional stacking integrated model,respectively.Then the average absolute percentage error distribution probability under each model is studied and the results are analyzed in depth.In summary,the improved stacking integrated model used in this paper can effectively improve the accuracy of each model in the prediction of reservoir parameters.A reliable prediction method with great potential could be used,and other wells in the study area can also be tested using it.Useful help for the future exploration and development of oil and gas fields can also be provided by the method.