Prediction of elastic parameters in gas reservoirs using ensemble approach

• Published in: Environmental earth sciences 2023-06, Vol.82 (11), p.269, Article 269
• Main Authors: Aghakhani Emamqeysi, Mohammad Reza, Fatehi Marji, Mohammad, Hashemizadeh, Abbas, Abdollahipour, Abolfazl, Sanei, Manouchehr
• Format: Article
• Language: English
• Subjects: Adaptive systems; Artificial neural networks; Biogeosciences; Correlation; Earth and Environmental Science; Earth Sciences; Empirical analysis; Environmental Science and Engineering; Fuzzy logic; Geochemistry; Geology; Geomechanics; Hydraulic fracturing; Hydrology/Water Resources; Laboratory tests; Machine learning; Mathematical models; Mechanical properties; Modelling; Modulus of elasticity; Neural networks; Original Article; Parameters; Poisson's ratio; Rocks; Root-mean-square errors; Stability analysis; Support vector machines; Terrestrial Pollution
• ISSN: 1866-6280; 1866-6299
• DOI: 10.1007/s12665-023-10958-4

Determination of the rock elastic parameters is essential in geomechanical studies. Among the elastic parameters, Young’s modulus (YM) and Poisson’s ratio (PR) have many applications in wellbore stability analysis, hydraulic fracturing, casing design, and sand production. In this study, the machine learning methods, including adaptive Neuro-Fuzzy inference system (ANFIS), artificial neural network (ANN), and support vector machine (SVM) are used to predict the rock elastic parameters. Using these models requires measuring the static elastic parameters, so 34 laboratory tests are used to develop empirical correlations between static elastic and dynamic elastic parameters. Then, the static elastic parameters at all logged intervals are calculated by applying the suggested empirical correlations. To demonstrate the capabilities of the ANFIS, ANN, and SVM methods, DT, RHOB, and NPHI data are used as inputs, and YM and PR data are used as outputs. The performance of single models can be enhanced using ensemble models, such as simple averaging ensemble (SAE), weighted averaging ensemble (WAE), and neural network ensemble (NNE). The results of the single models showed that ANN models performed better overall than other single models. The results also showed that ensemble models predicted elastic parameters better than single models. This shows that NNE model with R 2 of 0.998 and 0.993, MAPE error values of 0.0041 and 0.0010, and RMSE error values of 0.58 and 0.0029 for the training data of YM and PR is more accurate and reliable than SAE, WAE and single models.