Abstract:

The accuracy of evaluating total organic carbon in shale reservoirs is limited by using conventional logging curves because of
their insufficient generalization ability and requirement of a large number of samples. In view of these problems, neural network algorithm was improved to improve the prediction ability of the model. The cellular neural network structure was optimized by using a fuzzy system to enhance its logical reasoning ability and to improve its sensitivity to fuzzy data. The rain forest algorithm, which can effectively avoid the virtual collision, was selected, and its defect of slow convergence in the late learning was overcome. The initial weight value and threshold value of the network were optimized by the improved rain forest optimization algorithm to prevent the network from resulting in local minimum, which can improve the accuracy and generalization ability of the model. Based on the analysis of the physical meaning of the characteristic curve, the density log curves and the natural gamma ray spectrum logging curves were chosen as the input to the network and the total organic carbon content was used as the output. Through the network learning of 70 core samples and the prediction of 26 core samples, the role of the improved rain forest algorithm and fuzzy logic is proved. The superiority of the new network model is demonstrated. The result shows that the relative regressional error of the new model is reduced from 23.189% to 17.185%, and the relative prediction error is reduced from 52.421% to 15.158%, which means that the prediction is in accordance with the real situation of formation. From the above, we learn that the new model has better learning ability and generalization ability. The new model is more suitable for logging evaluation of total organic matter content in shale reservoirs.

Key words: shale gas；TOC；fuzzy neural network；improved rainforest algorithm；generalization