Estimation of Dead Oil Viscosity Utilizing Physics Based Correlative Principles and Predictive Machine Learning Techniques.

Abstract

Abstract Dead oil viscosity is one of the most unreliable properties to predict with classical black oil correlations. This results mostly from the large effect that oil type has on viscosity. Two dead oil samples with identical API and T can have even an order of magnitude difference in viscosity [9]. In this work, we tried to limit this spread to a certain degree by incorporating a parameter such as MW to capture additional information for the character of the oil. Limitations of the classical black oil correlations became even more prominent when wide-spectrum of viscosity values coupled with wide range of temperatures are considered. Given the constraints of limited input variables, the problem becomes particularly challenging for heavy-extra heavy oils with high asphaltene content [35], where prediction errors could easily be as high as a couple of log-cycles. Even though there are several viscosity correlations available in the industry [1]-[4], [7], [9], [26], [30], [31], [34], [37], [39], [42] most of those correlations are only applicable to the oil samples belonging to specific geographical regions and/or for structurally similar oils, because of the inherent bias in the training datasets used in the development of correlations. Therefore, they are predictively valid for a relatively narrow range of oils and/or viscosity. In this work, we considered a very wide range of oils (6° API to 50° API). Therefore, producing two easy to use viscosity correlations for API gravities above and below 20° API that can readily predict the viscosity at any desired temperature within an extended temperature range (15° C to 160° C). In addition, the two sets of correlations were kept continuous at the switching point (20 API) so that they will have smooth transition from one branch to another. While the range of fluid properties is considered to be very wide, we were able to keep the input parameters to a minimum in terms of defining the character of the fluid (molecular weight and specific gravity). We demonstrated that the proposed correlation performs much better than the leading correlations with the similar input proxies published in the literature for wide range of viscosities (0.5 cp to 860, 000 cp). The use cases for the proposed correlation can be divided into three parts: 1) Prediction of the dead oil viscosity with limited input data, 2) use of limited in-hand viscosity data to generate viscosities for the conditions that are hard to perform accurate experimentation or simply not in hand for various reasons (for example not having the physical sample to perform such additional experiments, or just not having the time and the sources such as in the case of, screening existing assets for thermal recovery where typically viscosity is available at reservoir temperature, not at steamflood temperatures) and 3) it can be used to check the consistency and the quality of the existing data. Therefore, we have also checked the predictive performance of the proposed correlation in temperature domain if limited viscosity information is available for a particular fluid at a different temperature. With such information (if one reference viscosity is available at some specific temperature) the accuracy of the correlation can be further improved. In addition to classical correlation development efforts using known but limited physical control parameters, we have also attempted to model the viscosity with various machine learning methods (such as Artificial Neural Network (ANN) [20], K-Nearest Neighbour (KNN) [6] and Kernel based Support Vector Machine (KSVM)) [18] [38] and compared the outcome with respect to each other and as well as against the proposed correlation. Our correlation showed better predictive capabilities (based on calculated statistical parameters and cross-plots) when compared with other leading viscosity [4], [31], [37], [39], [42] correlations as well as with relevant supervised machine learning regression principles such as ANN, KNN and KSVM. The subject correlation also helps in improving the accuracy as well guiding the performance of these otherwise "blackbox" machine learning principles as it can fill the gaps in the data especially in the context of extending the tuned viscosities based on a single point measurement in temperature domain. Furthermore, we also explain how it can be combined with Sinha et. al [35] relative viscosity correlation to include the impact of asphaltene concentration, for estimation of vertical or areal viscosity variations which can also ultimately help to improve the mobility cut-off predictions of the asphaltene/tar mat zones or heavier fluids.