A New Method to Improve the Prediction Accuracy on TOC of Source Rock by BP Neural Network Model

Abstract

The total organic carbon (TOC) has been one of the most important parameters of source rocks evaluation. Many great achievements were made in the study of TOC by using logging data which contributes to the occurrence of the calculation methods that were proposed and wildly employed in practice and production, such as △lgR technique, multivariate regression model, etc. However, problems like low accuracy due to the insufficient well logging parameters input (△lgR technique only involves acoustic transit time log and resistivity log) and unreasonably dealing with the nonlinear relationship between different well logging parameters and TOC, and poor universality of the models because of different geological environment still exist. So on the basis of the five most widely used models, including △lgR technique, CARBOLOG model, U-△lgR binary parameters model, multivariate regression and stepwise regression models, the paper introduces the BP neural network model. By taking the distinctive advantages of parallel processing of information, nonlinear approximation, self-learning and self-adaption of BP neural network model, can the limitations of the existing models be overcome.

Moreover, in order to evaluate the differences between the BP neural network model and the five models, the TOC of the source rocks was predicted in Ordos Basin and Bohaiwan Basin respectively, China. The results can be concluded as follows: (1)The error of BP neural network model is generally less than 10% (Lower than 25% in the original five models) by introducing the seven logging parameters of AC (acoustic transit time log), DEN (density log), GR (gamma-ray log), SP (natural potential log), U (U in the data of gamma ray spectrometry log), Th (Th in the data of gamma ray spectrometry log)and K (K in the data of gamma ray spectrometry log)(the number of the reference parameters of the original five models is less than five)to establish a nonlinear signal processing network; (2)The average correlation index (R²) of predicted and measured TOC values is higher than 0.8 in both basins (Higher than 0.6 in the original five models).

The research illustrates that the BP neural network model has theoretically broken through the limitations of the existing five models, and greatly improved the accuracy of TOC prediction in practice, which provides a solid and reliable geological basis for the prediction of favorable source rocks in Craton and Rift basins.

AAPG Datapages/Search and Discovery Article #90323 ©2018 AAPG Annual Convention and Exhibition, Salt Lake City, Utah, May 20-23, 2018