Abstract

An innovative and more accurate method for formation breakdown pressure (FBP) prediction based on self-learning artificial neural networks (ANNs) is presented. Results are compared to industry wide accepted deterministic FBP prediction models. ANN model validity is proven to meet FBP prediction within a 10% error range demonstrating that ANN models are outperforming the eleven deterministic prediction models investigated.

Using field data from four different countries neural network models were trained including automatic prevention of over fitting and over training as well as avoiding to a great extent trapping in local error minima. Since field data are very heterogeneous, statistics about the model performance was controlled and validated for quality assurance via multifold cross validation.

The study is based on125 plus 14 (for verification purposes) data sets from fracture work executed with FBPs ranging from 4,333 to 16,707 psi. In a first step the data were applied to deterministic models used in the industry to investigate, evaluate and prove their validity. The results were compared and analysed according to an error analysis framework identifying their accuracy and precision. Six error analysis parameters were determined and the individual models ranked accordingly. Although some of the deterministic models perform well in specific areas, it was not possible to individuate a model which is valid for all investigated data sets meeting the error criteria of ± 10%. In the second step the same data were used to train neural network models. The results were evaluated, refined and QA/QCed for their robustness. The error analysis based on multifold cross validation proved that the proposed ANN models and methods provide consistent and robust results in nearly all cases within a ±10% error range from the actual measured value in all types of reservoir rock.

The novelty of the proposed method is its high accuracy and robustness in heterogeneous conditions. It was successfully applied to large FBP ranges, different locations and all types of reservoir formations. Because of matured QC tools to prove accuracy and validity of the proposed model, it can be easily retrained to handle new conditions providing results according to the set error criteria when additional real FBP data are available. For specific fields this can lead to a more restrict predicition criteria range. At the same time, the model can be used in case of incomplete data.

You can access this article if you purchase or spend a download.