An AI Based Water Cut Sensor for Oil Wells

Abstract

Abstract In the oil and gas industry, converting data to actionable information by having an accurate measurement of the well real time data flow parameters is crucial. Such measurement process is fundamental for decision making by engineers; therefore, it requires state of the art gadgets to be installed in the field to send accurate real time data to the engineer desktop to monitor, analyze and make informed decisions. Water cut is one of the important parameters of well rate testing process where it requires periodic calibrations of the metering equipment to ensure high accuracy throughout the well operating life. This calibration task of the equipment is part of the maintenance program which requires to shut-down the well and lose its potential during these maintenance activities. This paper shows how to capitalize on real-time data from MPFM and ESP to predict the water using artificial machine learning models. the estimated water cut can be used to ensure the reliability of the existing vesical meters and optimize the required maintenance frequency which will minimize the well shut-down time. Candidates producers were selected to use and test several machine learning algorithms to evaluate their performance across different conditions and states. The study revealed good match between actual and estimated water cut figures which introduces a reliable redundancy tool to ensure equipment reliability and minimize the oil locked potential associated with maintenance activities. Evaluation of the efficacy of various machine learning algorithms on our dataset was conducted, we trained and tested four models: random forest, linear regression, MLP, and KNN models. Our results demonstrate that the random forest algorithm outperformed the other three models, achieving a remarkably high R-squared value of 0.9731, indicating that the model accounts for 97.31% of the variance in the response variable. We also computed other metrics to assess model performance, including the mean absolute error (MAE) and mean squared error (MSE), which confirmed that the random forest model had the lowest error rates. These results suggest that the random forest algorithm is the most suitable for our dataset and may be particularly effective for predicting our response variable. Additionally, a sensitivity analysis was performed for the random forest model This study provides a new solution for the oil and gas industry by designing a soft water cut sensor based on random forest model. The implementation of this sensor in the field can lead to improved decision making and optimization of field production strategies by accurately predicting water cut in real-time.