Research Of Reservoir Parameters Prediction Method Based On Machine Learning

Seismic attribute technology is an effective method for reservoir prediction.However,as an important part of oil and gas exploration and development,reservoir prediction requires higher prediction accuracy.With the deepening of exploration and development,exploration targets have gradually become smaller,deeper and more complex.From structural oil and gas reservoirs to subtle oil and gas reservoirs,traditional prediction methods can no longer meet current needs.Therefore,there is an urgent need for new technologies and methods to mine the information in seismic attributes,establish a nonlinear relationship between seismic attributes and reservoir parameters,and further meet the requirements of the exploration and development of complex reservoir oil and gas fields.In recent years,machine learning has been applied in many fields and achieved obvious results,which has promoted the application of this method in geophysics.Therefore,many scholars have done a lot of research on seismic attribute analysis by using machine learning methods,and achieved relatively ideal results in the field of oil and gas reservoir prediction.Among them,artificial neural networks(ANN)has strong self-learning and adaptive performance,and can fit arbitrarily complex nonlinear relationships well;random forest(RF)can well overcome overfitting,and has strong tolerance to outliers and fast training;support vector machines(SVM)can well solve the problem of few samples,nonlinearity,high dimensionality and local optimal value.Therefore,this paper used above three machine learning methods to build a model between seismic attributes and reservoir thickness to achieve the thickness prediction of the unknown area in the study area,and compared them with the prediction effect of MLR.Based on the fine interpretation of the glutenite reservoirs in the study area,this paper found that there were fewer wellheads and fewer training samples can be used.First,the traditional MLR was used to predict the thickness of the reservoir,and the prediction effects of different types of attributes were studied,and compared with the prediction effects of the ANN model.Through the error comparison analysis on the test samples,it was found that the performance of the ANN based on the nonlinear model was better.Because the ANN prediction results are unstable,phase control constraints were added to further improve the prediction accuracy.Then based on two different dimensionality reduction optimization algorithms,the RF model was used for reservoir thickness prediction,and the prediction effects of the two methods were compared and analyzed.Finally,for the two key model parameters of SVM,based on the idea of Cross Validation(CV),grid search and Fruit Fly Optimization Algorithm(FOA)were used to optimize the optimization,and a good prediction effect has been achieved.Research showed that machine learning algorithms based on nonlinear models have higher accuracy than MLR,but the accuracy of conventional prediction models could not meet the requirements.The accuracy can be further improved by improving the three models.Through comparison,the FOA-SVM model has the best performance.It was used to predict the porosity and oil saturation of the study area,and achieved ideal prediction results.In addition,based on the Visual Studio2017 platform,the seismic reservoir parameter prediction software was developed using C# language.The data preprocessing,attribute optimization and reservoir parameter prediction methods introduced in this article were packaged to facilitate the use and promotion of researchers.