|Theme||Visible||Selectable||Appearance||Zoom Range (now: 0)|
In this study, parameter estimation of an Autonomous Surface Vehicle using neural networks is tried. Parameters were estimated by using simulation result of trained neural network. With this estimated parameter, linear system was constructed and linear simulation was performed. By comparing simulation data and estimated result, this paper shows the validity of parameter estimation method using neural networks.
Accurate estimation of formation parameters from well logs is essential to quantitative log analysis. The conventional log analysis techniques are largely dependent on the initial assumptions about log analysis parameters. However, without any assumptions and mathematical models, neural network (NN) based formation parameter estimation approaches can quickly estimate formation parameters directly from well logs after successful training. In this paper, a four-layer feedforward neural network, and a neural model of principal component analysis and linear regression are used to estimate porosity from well logs. And a model-based inversion of well logs and a principal component regression algorithm are used for comparison with the above two neural approaches. Several examples from a Chinese oilfield demonstrate the NNs-based approaches' advantages over the conventional methods in estimating formation parameters from well logs. Even though the results show that the estimates of formation parameters from the four-layer NN approach are in the best agreement with the core data, when applying to different wells with complex geological structures, this four-layer NN approach often fails to give satisfactory and consistent results. In consideration of such limitation of a single net, the paper presents an integrated approach to solve this problem.
Quantitative log analysis is conventionally based on a series of mathematical formulas, or models derived from many log analysts' experience. These formulas simply use simplified equations to express complex correlative relationships between real log readings and formation parameters, arising from the high heterogeneity and anisotropy of the earth medium. In complex reservoir cases, these formulas often fail to produce accurate log analysis results with great inconsistentency with some of the initial assumptions about log analysis parameters. Applications of neural networks in quantitative log analysis, with NNs, adaptability to and learning of various rock types and borehole environments as well as multi-layer NNs' capability of expressing arbitrary complexity of input data presented to NNs, can evidently reduce dependence of log analysis on the assumptions about various log analysis parameters. Without any assumptions and any mathematical models, NN-based formation parameter estimation approaches can quickly estimate formation parameters directly from well logs after successful training with real log measurements and the corresponding core data in the same well. In this paper, a four-layer feedforward neural network, and a neural model of principal component analysis and linear regression (PCALR) are used to estimate porosity from well logs, and a model-based inversion of well logs and a principal component regression algorithm are used for comparison with the above neural approaches. After detailed analysis and comparison of these four methods, the paper presents the differences among these methods and their limitations. Several examples from a Chinese oilfield demonstrate NNs-based approaches' advantages over the conventional methods in estimating porosity from well logs. The comparison shows that of all the methods mentioned above, the four-layer NN approach has produced porosity estimates in a best agreement with core data. But when applying for different wells in a region with complex subsurface geological structures, even the four-layer NN approach can not give porosity estimates satisfactory and consistent enough.
Al-AbdulJabbar, Ahmad (King Fahd University of Petroleum & Minerals) | Al-Azani, Khaled (King Fahd University of Petroleum & Minerals) | Elkatatny, Salaheldin (King Fahd University of Petroleum & Minerals)
Porosity is one of the most important properties to be determined for evaluating hydrocarbon reservoirs. It represents the voids and empty volume inside the rock. This property is mostly obtained from well logs and/or laboratory experiments on core plugs or drilled cuttings. Despite the accuracy in the porosity values provided by these techniques, these methods are costly and time consuming. There is a need to relate the rock porosity to the drilling parameters since drilling process provides the initial insight to the formation. The use of artificial intelligence (AI) in drilling applications is a game changer since most of the unknown parameters are accounted during the modeling process. The objective of this paper is to implement an artificial neural network (ANN) technique to predict the porosity in the reservoir section from the drilling parameters. The data used to build the ANN model are based on real field data (2,800 data points) that were obtained from two horizontal wells (i.e. Well A and Well B). The data from Well A were used to train and test the ANN model with a training/ testing ratio of 70:30. More than 30 sensitivity analyses were performed to select the optimum ANN model’s design parameters. Well B data were used to validate the developed ANN model. The obtained results showed that ANNs can be used effectively to predict the porosity from the drilling parameters in the reservoir section with an average correlation coefficient of approximately 0.96 and a root mean square error (RMSE) of almost 0.018. The best ANN parameter combination was with two layers, 30 neurons per layer with Levenberg-Marquardt training function and tan-sigmoid as the transfer function. The validation process confirmed that the ANN porosity model was able to predict the porosity of Well B with a correlation coefficient of 0.907 and an RMSE of 0.035.
Multisource and multiscale modeling of formation permeability is a crucial step in overall reservoir characterization. Thus, it is important to find out an efficient algorithm to accurately model permeability given well logs data. In this paper, an integrated procedure was adopted for accurate Lithofacies classification prediction to be incorporated with well log attributes into core permeability. Probabilistic Neural Networks and Generalized Boosted Regression Models were adopted for Efficient Lithofacies Classifications and Formation Permeability Estimation, respectively.
The Probabilistic Neural Networks (PNN) is an implementation of a statistical algorithm called kernel discriminant analysis in which the operations are organized into a multi-layered feedforward network with four layers: Input, Pattern, Summation, and Output layers. It was used to model Lithofacies sequences in order to predict discrete lithofacies distribution at missing intervals. Then, Generalized Boosted Regression Modeling (GBM) was used as a to build a nonlinear relationship between core and log data. GBM is a recent data mining technique that has shown considerable success in predictive accuracy as it maintains a monotonic relationship between the response and each predictor.
The well log interpretations that were considered for Lithofacies classification and permeability modeling are neutron porosity, shale volume, and water saturation as function of depth; however, the measured discrete lithofacies types are Sand, Shaly Sand, and Shale. Firstly, the Probabilistic Neural Networks was adopted for modeling and prediction the discrete Lithofacies distribution at missing intervals. The classified Lithofacies were considered as a discrete independent variable in core permeability modeling in order to provide different model fits given each Lithofacies type to capture the permeability variation. Then, GBM was applied to build the statistical modeling and create the relationship between core permeability and the explanatory variables of well logs and Lithofacies. In GBM results, Root Mean Square Prediction Error (RMSPE) and adjusted R-square have incredible positive values, as there was an excellent matching between the measured and predicted core permeability. The GBM model has led to overcome the multicollinearity that was available between one pair of the predictors. All the multivariate statistics analyses of Lithofacies classification and permeability modeling with results visualizations were done through R, the most powerful open-source statistical computing languages.
Based on the same dataset, the PNN Lithofacies algorithm is the best classification approach as the total percent correct of the predicted discrete Lithofacies has exceeded 97.5% in comparison with other methods such as Linear Discriminant Analysis and Support Vector Machine. In addition, the RMSPE and Adjusted R-square obtained by GBM are much better than linear regression methods and Generalized Additive Models that have been applied on the same data as well.