首页|Evaluation of different machine learning frameworks to predict CNL-FDC-PEF logs via hyperparameters optimization and feature selection
Evaluation of different machine learning frameworks to predict CNL-FDC-PEF logs via hyperparameters optimization and feature selection
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Although being expensive and time-consuming, petroleum industry still is highly reliant on well logging for data acquisition. However, with advancements in data science and AI, methods are being sought to reduce such dependency. In this study, several important well logs, CNL, FDC and PEF from ten wells are predicted based on ML models such as multilinear regression, DNN, DT, RT, GBoost, k-NN, and XGBoost. Before applying these models, depth matching, bad hole correction, de-spiking, and preprocessing of the data, including normalization, are carried out. Three statistical metrics, R2, RMSE, and PAP, are applied to evaluate the models' performance. Results showed that RF, k-NN, and XGBoost are superior to others. While hyperparameters of the best models are optimized by GA, results from optimization demonstrate that each models' performance in predicting different logs can be improved by at least 1.5%. Furthermore, these models are evaluated for feature selection, done by GA, presenting that preserving all data in proposed models will improve the performance to the highest degree while reducing the number of features will deteriorate their performance. Comparison of performance measures for different combinations exhibited that the prediction of CNL-FDC-PEF logs with fewer inputs could be possible with relatively satisfactory outcomes. This study was innovative in incorporating all possible steps that can constitute a comprehensive ML model to improve well log data prediction. Moreover, it confirms that such methods will benefit us by reducing operational costs, time, and risks of tool failure in the wellbore by running a fewer number of well logs when data acquisition can be replaced by a comprehensive ML predictive model.
Drilling Engineering Chemical & Petroleum Engineering Department, Sharif University of Technology, Tehran, Iran
Department of Chemical Engineering, Engineering School, University of Sao Paulo (USP), Sao Paulo, Brazil
Key Laboratory of Continental Shale Hydrocarbon Accumulation and Efficient Development, Ministry of Education, Northeast Petroleum University, Daqing, 163318, China
NSTL
