This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 220686, “Real-Time Well-Status Prediction Using Artificial-Intelligence Techniques for Accurate Rate Allocation,” by Mohammad S. Al-Kadem, SPE, Abdulrahman Alajmi, and Najmul Ansari, SPE, Saudi Aramco, et al. The paper has not been peer reviewed.
Knowing well-operating conditions can help with accurate rate allocation; however, several factors govern well status, such as wellhead or downhole temperature and pressures. In this study, artificial-intelligence (AI) techniques were used to estimate and predict well status using a combination of surface and subsurface parameters in offshore areas. Four machine-learning (ML) algorithms were used to estimate and then predict well-operating status.
The accuracy of predicting well-operating status relies on the appropriate selection of suitable and robust parameters directly related to that status. It is preferable to limit the selection of the parameters to the most important ones to avoid an unnecessary increase in the complexity of the model. Several well parameters are available to be used in the ML model are obtained from surface and subsurface gauges installed on the well. In general, wells equipped with surface gauges are more popular than wells equipped with subsurface gauges; therefore, it is recommended to use surface parameters only to generate a model that can be generalized on most wells. The parameters below are detailed in the complete paper:
- Choke-valve upstream pressure/downstream pressure
- Choke-valve upstream temperature/downstream temperature
- Choke-valve position
The complete paper studies seven of the more-prominent ML algorithms—linear regression, logistic regression, random forest (RF), decision tree (DecT), support vector machines (SVMs), gradient-boosting machine (GBM), and artificial neural networks (ANNs)—by determining their definitions, their main properties, and the meanings of these properties; determining the degree of accuracy in each algorithm; and exploring their advantages and disadvantages. While all seven are detailed in the complete paper, RF is detailed in this synopsis.
RF, an ensemble learning algorithm, combines multiple decision trees to make predictions. It operates by constructing multiple trees from randomly sampled subsets of the training data and features and outputting the class that is the mode of the classes (classification) or the mean prediction (regression) of the individual trees. This ensemble approach makes RF robust to noise and outliers and enables it to capture nonlinear and complex relationships between input and target variables. One of the key advantages of RF is its ability to handle high-dimensional data and many input variables, making it suitable for applications such as genomic data analysis and image recognition. It also provides feature-importance measures, which can be used for feature selection and gaining insights into the most influential variables. RF can be computationally expensive to train, however, especially for large data sets, and may require careful tuning of hyperparameters.
