Desenvolvimento de um sensor virtual para estimativa da viscosidade aparente de fluidos de perfuração à base de água

Abstract

Oil exploration in offshore fields located in deep and ultra deep waters have been one of the challenges in oil production. To facilitate the operation, numerous studies are adding technology to the drilling fluid in order to meet the needs required during drilling of the geological formation. Thus, understanding and monitoring of the fluid properties are of vital importance for the exploration and production of oil. The objective of this work was to develop a soft sensor to estimate a rheological property of water based drilling fluids: the apparent viscosity. This sensor was based on the artificial neural network model with structure of multi-layer perceptron type. The initial stage of this work consisted in investigating the influence the additives in apparent viscosity in drilling fluid. For this purpose, full factorial design at three levels and with three replicates at the center point was carried, where the additives used were xantan gum, bentonite and barite. Once the temperature of the drilling fluid returning from the well to the surface is different from that injected into the drill string this variable was also considered in the planning in order to assess their influence on viscosity. Through this planning 84 water based drilling fluids were formulated; in which the apparent viscosity was determined in a viscometer (Fann 35A) at 300 RPM speed and at different temperatures. From the statistical analysis of results, it was found that all the variables considered had influence on apparent viscosity, and through the ANOVA test, analysis of variance, it was possible to obtain a regression model that takes into account the cross effects of each component in the response variable. The neural network used in this work is the type MLP, containing hyperbolic activation functions in neurons in the hidden layer and a linear output neuron. A total of it 750 and 267 experimental data were used for training and validating the neural network respectively, which were obtained by varying the temperature between 20oC and 60oC in each of the formulated 84 fluids in the planning stage of experiments. Several simulations were performed during the learning stage of the neural network by varying the number of neurons in the hidden layer in order to obtain the best architecture. It was found that the network with 6 neurons in the hidden layer had the best generalization ability and, therefore, the best predictions. The predictive ability of the selected neural network was compared to the statistical regression model using the mean square error and visual inspection test. It was found that the performance of the neural model was higher, presenting significantly lower prediction errors.