The Use of Regression and Classification Algorithms for Layer Productivity Prediction in Naturally Fractured Reservoirs

摘要

Three regression algorithms and three classification algorithms have been applied to forecast naturally fractured layer productivity. The three regression algorithms are the regression of support vector machine (R-SVM), the back-propagation neural network (BPNN), and the multiple regression analysis (MRA), while the three classification algorithms are the classification of support vector machine (C-SVM), the naive Bayesian (NBAY), and the Bayesian successive discrimination (BAYSD). In general, when all these six algorithms are usedto solve a real-world problem, they often produce different solution accuracies. When an algorithm is applied to a real-world problem, its solution accuracy is expressed with the total mean absolute relative residual for all samples, R(%). In this paper, three criteria have been proposed: 1) nonlinearity degree of a studied problem based on R(%) of MRA (weak if R(%)<10, moderate if 1030); 2) solution accuracy of a given algorithm application based on its R(%) (high if R(%)<10, moderate if 10< R(%)<30, and low if R(%)>30); and 3) results availability of a given algorithm application (applicable if R(%)<10, and inapplicable if R(%)>10). A case study of naturally fractured layer productivity at an oilfield in Sichuan Province of China has been used to validate the proposed approach. This case study consists of two problems: regression and classification. For the regression problem, .R-SVM, BPNN and MRA are inapplicable because their R(%) values are 4620, 44 and 5980, respectively/For the classification problem, however, C-SVM, NBAY and BAYSD are all applicable since their R(%) values are 0, 5.9 and 9.1, respectively. From the case study, it is concluded that: a) the three proposed criteria, and the rules of conversion from real number to integer number are practical; b) R(%) of MRA is. used to measure the nonlinearity degree of a studied problem, and thus MRA should be used as a first choice; c) for classification problems, the preferable algorithm is C-SVM, NBAY, or BAYSD if the problems nonlinearity is weak or moderate, and BAYSD can also serve as a promising dimension-reduction tool; d) for regression problems, the preferable algorithm is BPNN, but MRA can serve as a promising dimension-reduction tool only when the problems are linear; e) if BPNN is inapplicable for a regression problem, it is proposed to change the problem from regression to classification by reasonable conversion rules, then apply C-SVM, NBAY, or BAYSD; and f) comparing with C-SVM, BAYSD is conditionally better.