Abstract
From the perspective of wireline formation test (WFT), formation tightness reflects the "speed" of pressure buildup while the pressure test is being conducted. We usually define a pressure test point that has a very low pressure-buildup speed as a tight point. The mobility derived from this kind of pressure point is usually less than 0.01md/cP; otherwise, the pressure points will be defined as valid points with valid formation pressure and mobility. Formation tightness reflects the formation permeability information and can be an indicator to estimate the difficulty of the WFT pumping and sampling operation.
Mobility, as compared to permeability, reflects the dynamic supply capacity of the formation. A rapid and good mobility prediction based on petrophysical logging can not only directly provide valid formation productivity but can also evaluate the feasibility of the WFT and doing optimization work in advance.
Compared to a time-consuming and costly drillstem test (DST) operation, the WFT is the most efficient and cost-saving method to confirm hydrocarbon presence. However, the success rate of WFT sampling operations in the deep Kuqa formation is less than 50% overall, mostly due to the formation tightness exceeding the capability of the tools. Therefore, a rapid mobility evaluation is necessary to meet WFT feasibility analysis.
As companion work to a previous WFT optimization study(SPE-195932-MS), we further studied and discuss the machine learning for mobility prediction. In the previous study, we formed a mobility prediction workflow by doing a statistical analysis of more than 1000 pressure test points with several statistical mathematic methods, such as univariate linear regression (ULR), multivariate linear regression (MLR), neural network regression analysis (NNA), and decision tree classification analysis (DTA) methods.
In this paper, the methods and principles of machine learning are expounded. A series of machine learning methods were tested. The algorithms that are appropriate for these specific data set were selected. Includes DTA, discriminant analysis (DA), logistic regression, support vector machine (SVM), K-nearest neighbor (KNN) for formation tightness prediction and linear regression, DTA, SVM, Gaussian process regression SVM, random tree, neural network analysis for mobility prediction.
Contrastive analysis reveals that:
The SVM classifier has the best result over other methods for formation tightness probability prediction. Based on R squared and RMSE analysis, linear regression, GPR, and NNA delivered relatively good results compared with other mobility prediction methods.
An optimized data processing workflow was proposed, and it delivered a better result than the workflow proposed in SPE-195932-MS under the same training and testing dataset condition.
The comparison between measured mobility and predicted mobility results reveals that, in most situations, the predicted mobility and measured mobility matched very well with each other. WFT were conducted in newly drilled wells. Sampling success rate also achieved 100% in these wells by optimizing the WFT tool string and sampling stations selection in advance, and NPT is significantly reduced.
A comparative study of multiple linear regression, artificial neural network and support vector machine for the prediction of dissolved oxygen
