Prediction of the Least Principal Stresses Using Drilling Data: A Machine Learning Application

Estimating least principal stresses during drilling is essential for optimizing operations and preventing formation damage. Traditional methods rely on field tests and theoretical models that require additional time and resources. These approaches often fail to provide real-time data, which limits their usefulness in dynamic drilling environments. While injection rate, standpipe pressure, and other drilling parameters are continuously collected, their direct use in stress estimation has been limited. Prior research has shown that tectonic stress influences stress profiles but has not fully integrated this with machine learning techniques. The need for a cost-effective and timely method motivates the development of new predictive models. Existing models lack the accuracy and adaptability required for complex field conditions. This gap motivated the exploration of artificial neural networks as a potential solution. No prior work had resolved how to use drilling data directly for stress estimation without additional field tests.

This study aimed to develop a novel machine learning-based approach to estimate the least principal stresses during drilling. The goal was to use readily available drilling data to predict stresses without relying on costly field tests. The focus was on creating models that could provide real-time stress values during operations. The approach sought to reduce the need for leak-off tests and other intrusive methods. The study specifically targeted the prediction of minimum and maximum horizontal stresses. The motivation was to improve drilling efficiency and reduce operational risks. The use of artificial neural networks was chosen for its adaptability and accuracy. The ultimate aim was to provide a reliable and economical alternative to traditional stress estimation methods.

Artificial neural networks (ANN) were employed to create predictive models for stress estimation. The models used injection rate, standpipe pressure, weight on bit, torque, and rate of penetration as input variables. These parameters were selected based on their availability during drilling operations. The data were collected from a Middle Eastern field and preprocessed for statistical analysis. The dataset was divided into training and validation sets to ensure model accuracy. The ANN models were trained using a supervised learning approach with backpropagation. The performance of the models was evaluated using correlation coefficients and mean absolute average error. The results were compared against actual stress values to assess model reliability.

The ANN-based models demonstrated high accuracy in predicting minimum and maximum horizontal stresses. The correlation coefficient exceeded 0.90 for both stress models. The mean absolute average error was as low as 0.36% for the minimum stress model. The maximum error observed was 0.96% for the maximum stress model. The models showed strong consistency with actual stress values from the field data. Empirical equations were derived from the ANN models for practical use. These equations were validated using an unseen dataset from the same field. The validation results confirmed the models' robustness and generalizability. The developed approach outperformed traditional methods in terms of accuracy and cost-effectiveness.

The study demonstrated that artificial neural networks can accurately predict least principal stresses using drilling data. The developed models provided reliable stress estimates with minimal error. The use of readily available drilling parameters eliminated the need for additional field tests. The empirical equations derived from the models can be applied in real-time drilling operations. The results suggest that machine learning can enhance drilling efficiency and safety. The high correlation coefficients and low error values support the models' practical utility. The approach offers a cost-effective alternative to traditional stress estimation methods. The findings align with the study's objective of providing a novel and accurate solution for stress prediction.