| dc.description.abstract | Poor hole cleaning can be quite costly as it can lead to different operational problems such as mechanical stuck pipe, premature bit wear, slow drilling progress, formation fracturing, excessive drillstring torque and drag preventing from reaching the target in highly deviated/ ERD wells, difficulties during logging and cementing, and difficulties during casing landing. In the worst case, it can result in the complete loss of a well. However, transport of cuttings during drilling is a complex phenomenon due to presence of multiphase flow, and different operational parameters (such as flow rate, ROP, RPM, and temperature), cuttings and fluid parameters (such as density, cuttings size and fluid rheology) and geometrical parameters (such as annulus size, inclination, and eccentricity) influencing the transport of cuttings out of the hole. Conventional approach for modeling downhole cuttings concentration (Cc), an indicator of hole cleanliness, is, hence, challenging.
Different empirical models are developed by different researchers based on experimental data to estimate Cc as an alternative modeling approach. In the current study two empirical models are considered. The first empirical model termed Emp 1 is developed by Ahmed et al. [1] and the second one termed Emp 2 is by Rubiandini [2]. Emp 1 is developed using dimensional analysis, while Rubiandini has extended the empirical cuttings slip velocity model of Larsen et al. [3] to cover the entire hole inclination angle range from 0 to 90o using experimental data of Larsen et al. and Peden et al. [4]. Larsen’s cuttings concentration estimation model (incorporating viscosity correction factor) is combined with Rubiandini’s critical mud velocity model to estimate Cc. For the experimental data published by Ahmed et al. using the TUDRP test facility (nine experimental observations), Emp 2 provides Cc estimates with comparable accuracy to Emp 1 for medium and large annuli while Emp 1 outperforms Emp 2 for small annulus. However, Emp 1 is found to be inapplicable for vertical wells with 0o inclination angle and for non-rotating drillstring cases (turbo-drilling) while Emp 2 can provide negative cuttings concentration if critical mud velocity is less than the applied pump rate.
In the pursuit of developing a better model than empirical ones mentioned earlier, a computational intelligence (CI) method, fuzzy logic (FL), is chosen to develop a Mamdani type FL model to estimate Cc using experimental data. FL, resembling human reasoning, is capable of modeling non-linear functions of arbitrary complexity (and hence, has the ability to deal with complex phenomenon) and dealing with imprecision in data. Among the different FL approaches, Mamdani FL system is popular due to its interpretability. The first FL model (termed FL 1) developed in the current study is based on 509 experimental observations involving 11 independent test parameters each. These experimental observations provide 195 fuzzy rules. 25% of the collected dataset (excluding the test dataset) is used as cross-validation dataset and the rest is used as training dataset to avoid overfitting. The same test dataset published by Ahmed et al. is used for comparing FL 1 with Emp 1 and Emp 2. The comparison shows that FL 1 model performs better than Emp 1 and Emp 2 in all the three goodness of fit metric namely coefficient of determination (R2), root mean square error (RMSE) and mean absolute error (MAE).
FL 1 model, however, has several limitations revealed after performing exploratory data analysis (EDA). EDA is performed on six experimental datasets (702 experimental observations) in a two-fold manner - univariate and multivariate analysis. Univariate analysis shows the asymmetry in distribution for each experimental parameter justifying the choice of FL, a nonparametric modeling approach, as an alternative model development approach to estimate Cc. Multivariate analysis shows the interaction of the experimental parameters among themselves and their influence on downhole Cc using 6D scatter plots and correlation coefficients (Kendall’s ꚍ). EDA of the experimental data reveals the following major findings:
Smaller Cc in concentric vertical wells compared to concentric non-vertical wells.
Drilling fluid flow rate is a dominant operational parameter in vertical wellbore cleaning while string rotation (RPM) is dominant in non-vertical wellbore cleaning.
Little impact of RPM in concentric vertical well and negative eccentric deviated/highly deviated well cleaning. However, RPM together with drilling fluid flow rate provides better cleaning of nonvertical wells with positive eccentricity.
RPM has higher influence on cuttings transport in narrow annulus compared to that in wide annulus
Assuming drilling fluid of sufficient viscosity and drill string rotation are present, low viscous fluid under turbulent flow and high viscous fluid under laminar flow provide better hole cleaning. Further, Kendall’s ꚍ indicates apparent viscosity playing a more significant role in cleaning deviated wellbores compared to other inclinations for the current dataset.
Drilling fluid flow rate influences the transport of heavier cuttings and larger cuttings more while RPM has higher influence on the transport of lighter cuttings and smaller cuttings.
Better hole cleaning by heavier drilling fluids than that by lighter fluids.
The findings from EDA are used to develop a second FL model termed FL 2 using 702 experimental observations providing 265 fuzzy rules. FL 2 is further combined with genetic algorithm (GA) to optimize fuzzy rule weights which has led to the development of a hybrid CI model termed FL 3. The performance of FL 3 over two test datasets collected at two different geographical locations by two different research groups – TUDRP (same test dataset as previously mentioned) and SINTEF (an additional test dataset for comparing all the three FL models)- are used. Model comparison using the three aforementioned goodness of fit metrics shows that the hybrid CI model, FL 3, outperforms both FL 1 and FL 2 over both test datasets. R2 value of FL 3 over both test datasets is ‘very good’ following Moriasi et al.[5]. The difference of FL 3 Cc estimates and the measured Cc values is less than 1% for more than two thirds of the experimental data for both test datasets. The reasons behind FL 3 outperforming FL 1 and FL 2 are data randomization, modified fuzzy sets for eccentricity and drillstring rotation based on EDA, use of more experimental data and the optimization of fuzzy rule weights by GA.
The developed hybrid CI model, FL 3, can be used during planning phase to select the optimum well design parameters (operational, cuttings/fluid, and geometrical). It can also be used during the execution phase to determine whether the hole is cleaned adequately. FL 3 can be further modified incorporating other CI methods (such as artificial neural network) and using more experimental/field data. | en_US |
