Abstract
In this work we propose a pure data-driven framework to categorize oil producing wells from a pre-salt field based on the chemical compositions of the original formation water, without depending on physical-chemical equations or simulation models. Our unsupervised learning method is intended to group wells with similar reactivities, which could require the same management actions. Taking as a starting point some formation water composition analyses, mostly sampled from injectors, our first step is to spatially interpolate the concentrations of each of the ions, resulting in maps that can give us estimates of the values in the vicinity of the producing wells. Then, we apply dimensionality reduction methods, as the Principal Component Analysis (PCA) to get a clearer view of how those wells are distributed in terms of water composition, and if any kind of grouping is possible. We then use the clustering methods K-means, Fuzzy K-means and Self-Organizing Maps (SOM) for the last step in knowledge extraction. The work focuses on an exploratory analysis to detect trends and groupings of wells with similar characteristics. Observing which wells are contained in each group is therefore useful to drive our efforts in actions such as the correct application of scale inhibitors both in the reservoir and in the production tubing; dimensioning the sampling frequency of produced water for analysis of chemical composition and inspection of solids; validation or adjustment of the injection strategy, considering the compatibility of the injection water with each group's formation water and the consequences of their mixture. The novelty presented is how we are extracting knowledge from the available data (scarce history of breakthroughs, but with an abundance of laboratory analysis) and using it as a decision support tool in the management of pre-salt carbonate reservoirs.
