Kick Prediction Method Based on Artificial Neural Network Model

Abstract

Kick is one of the most important drilling problems, and because its occurrence makes drilling engineering extremely complex, it is essential to predict the possibility of kick as soon as possible. In this study, k-means clustering was combined with four artificial neural networks: regularized RBFNN, generalized RBFNN, GRNN, and PNN, to estimate the kick risk. To reduce data redundancy and normalize the drilling data, which contain kick conditions, k-means clustering was introduced. The output layer weights were then determined using a brute-force search with different Gaussian function widths, resulting in a series of artificial neural networks composed of different clustering samples and different Gaussian function widths. The results showed that the prediction accuracy of regularized RBFNN + k-means model was the highest, that of the GRNN + k-means model was the lowest. The kick prediction accuracy for regularized RBFNN, generalized RBFNN, GRNN, and PNN were 75.90%, 65.20%, 51.70%, and 70.16%, respectively. This method can be used to enhance the speed and accuracy of kick risk prediction in the field while facilitating the use of and advances in risk warning technology for deep and high-temperature and high-pressure wells.

Keywords

• kick
• neural network
• k-means clustering
• normalized RBFNN
• data learning and training