Hybrid machine learning algorithms to enhance lost-circulation prediction and management in the Marun oil field

Abstract

Drilling fluid loss of circulation is a challenging issue to resolve for many oil and gas wells as drilling progresses. It imposes enormous expenses on drilling industry. One of the common practices to solve this problem or make it less severe is introducing various loss circulation materials (LCM) into the circulating drilling fluid. The ability to predict the amount of mud loss before drilling a particular formation would provide engineers with invaluable information to assist them in selecting appropriate LCM properties such as particle size distribution. In this study, a number of heuristic search algorithms including genetic algorithm (GA), particle swarm size (PSO), and cuckoo search algorithm (COA) are combined with two machine-learning techniques namely multilayer perception (MLP) neural network and least square support vector machine (LSSVM) to present different hybrid algorithms in prediction of lost circulation. The established database consists of 2820 cases (datasets) extracted from available drilling reports of 305 drilled wells in the Marun oil field. These datasets are associated with drilling operation variables, formation specifications, and drilling mud characteristics which totally include 18 input variables and 1 output. Pre-processing of data involves filtering with Savitzky-Golay (SG) filter and ranking features using wrapper method. Results show that hybrid intelligent models are highly capable of predicting lost circulation before drilling a certain formation. Furthermore, LSSVM-COA and LSSVM-PSO achieve coefficients of determination of 0.9424 and 0.9391, respectively, making them the two most accurate models. GA cannot match the prediction performance of the other two algorithms in enhancing machine-learning techniques for this purpose. MLP-GA with a coefficient of determination (R2) and RMSE of 0.9304 and 25.26 (bbl/hr), respectively, yields the weakest prediction performance. Hybrid intelligent models, coupling optimizers and machine learning algorithms show themselves to be superior to applying standalone machine-learning methodologies for predicting loss of circulation.