| Abstrakt: |
A major cause of some of serious issues encountered in a drilling project, including wellbore instability, formation damage, and drilling string stuck e which are known to increase non-productive time (NPT) and hence the drilling cost e is what we know as mud loss. The mud loss can be prevented or at least significantly reduced by taking proper measures beforehand provided the position and intensity of such loss can be properly predicted using an accurate predictor model. Accordingly, in this study, we used the convolutional neural network (CNN) and hybridized forms of multilayer extreme learning machine (MELM) and least square support vector machine (LSSVM) with the Cuckoo optimization algorithm (COA), particle swarm optimization (PSO), and genetic algorithm (GA) for modeling the mud loss rate based on drilling data, mud properties, and geological information of 305 drilling wells penetrating the Marun Oilfield. For this purpose, we began by a pre-processing step to attenuate the effect of noise using the Savitzky-Golay method. The whole set of available data was divided into the modeling (including 2300 data points) and the validation (including 483 data points) subsets. Next, the second generation of the non-dominated sorting genetic algorithm (NSGA-II) was applied to the modeling data to identify the most significant features for estimating the mud loss. The results showed that the prediction accuracy increased with the number of selected features, but the increase became negligible when the number of selected features exceeded 9. Accordingly, the following 9 features were selected as input to the intelligent algorithms (IAs): pump pressure, mud weight, fracture pressure, pore pressure, depth, gel 10 min/ gel 10 s, fan 600/fan 300, flowrate, and formation type. Application of the hybrid algorithms and simple forms of LSSVM and CNN to the training data (80% of the modeling data, i.e. 1840 data points) showed that all of the models tend to underestimate the mud loss at higher mud loss rates, although the CNN exhibited lower underestimation levels. Error analysis on different models showed that the CNN provided for a significantly higher degree of accuracy, as compared to other models. The more accurate outputs of the hybrid LSSVM model than those of the simple LSSVM indicated the large potentials of metaheuristic algorithms for achieving optimal solutions. The lower error levels obtained with the CNN model in the testing phase highlighted the excellent generalizability of this model for unseen data. The more accurate predictions obtained with this model, rather than the other models, in the validation phase further proved this latter finding. Therefore, application of this method to other wells in the same field is highly recommended. [ABSTRACT FROM AUTHOR] |
