Predição de Perfis Geofísicos de Poços por Classificador KNN

There is a large diversity of well log curves but, in practice, its availability is always limited. Thus, sometimes a desired log curve is not available for a given application. However, well logs are in some level interdependent, so, it is possible to use some well logs to estimate another missing curve. In this work the multivariate statistical method KNN (K-nearest neighbors) is used to estimate a basic suite of well logs (GR, RHOB, NPHI, DT and ILD) from Namorado Field, offshore Brazil. KNN is adopted due to its simple implementation, low computer cost and high resolution. A training data set was composed by random choice of 30% of standardized well log curves from 12 vertical wells. The remaining data was used for well log prediction. All wells have the complete suite of logs, but in order to check the performance of prediction, each time one log curve was removed from data base, a synthetic one was estimated and compared to the original curve, furnishing an average prediction error. KNN was a suitable method for synthetic curve estimation of the majority of well logs, with estimated curves well correlated to real curves in terms of curve shape, well log values and resolution level. Nevertheless, the same prediction performance was not achieved for all log curves. For ILD curve the general prediction error was 186.9%, an unacceptable high value, meanwhile the prediction error was clearly satisfactory for GR (13.2%), NPHI (12.6%), DT (4.1%) and RHOB (1.5%). Introducao