Precision Forecasting: Machine Learning for Liquid Flow Rates through Wellhead Chokes

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Mar 27 2024

In an article recently published in the journal Nature, researchers focused on predicting fluid flow rates in production wells using two approaches: data-driven modeling and a developed correlation based on wellhead pressure (Pwh), gas-to-liquid ratio (GLR), and choke size.

The authors compared the performance of these approaches with existing correlations and evaluated the sensitivity of input variables. Results indicated superior performance of the proposed models highlighting the significance of accurate flow rate prediction in hydrocarbon recovery.

Background

The control of fluid flow rates in oil and gas production, particularly through wellhead chokes, is vital for regulating production rates and ensuring optimal hydrocarbon recovery. Wellhead chokes, either fixed or adjustable, play a crucial role in managing production rates and mitigating issues such as formation damage and gas coning.

Understanding fluid flow behavior through chokes, especially in two-phase flow scenarios, is essential for effective production management. Previous studies have proposed empirical correlations and theoretical models to predict flow rates through wellhead chokes, but many suffer from limitations such as high error rates and a lack of comprehensive modeling approaches.

The present paper addressed these gaps by proposing two novel approaches to predict liquid flow rates through wellhead chokes. Firstly, it introduced machine learning-based models, including adaptive boosting support vector regression (AdaBoost-SVR), multi-layer perceptron (MLP), radial basis function (RBF), and multivariate adaptive regression spline (MARS) algorithms, to accurately forecast flow rates. These models leveraged a comprehensive dataset comprising crucial variables such as Pwh, choke size, and (GLR).

Secondly, the authors developed a new empirical relationship for predicting flow rates, aiming to outperform existing correlations. By applying statistical evaluation, sensitivity analysis, and the leverage method, the researchers provided insights into the relative impacts of input variables on flow rates and identified outlier data points. Overall, this research significantly contributed to enhancing the accuracy and reliability of predicting liquid flow rates through wellhead chokes, addressing the limitations of previous empirical and theoretical models.

Model development

The study focused on predicting liquid flow rates through wellhead chokes using a comprehensive dataset of 565 data points collected from various sources. The key parameters influencing liquid flow rates, namely Pwh, choke diameter, and GLR, were identified as crucial inputs for the predictive models. Statistical analysis of the dataset revealed a wide range of liquid flow rates, with parameters such as mean, minimum, maximum, kurtosis, skewness, and standard deviation (SD) examined to understand the distribution of the data.

To develop accurate predictive models, the study employed several machine learning algorithms, including MLP, RBF, AdaBoost-SVR, and MARS. These algorithms offered diverse approaches to capture the complex relationships between input parameters and liquid flow rates. For instance, MLP utilized multiple layers of neurons to represent non-linear mappings, while RBF focused on specific distance-based processing units. AdaBoost-SVR employed a collective learning method to combine multiple classifiers' outputs for improved accuracy, and MARS explored non-linear relationships through basis functions.

Additionally, the generalized reduced gradient (GRG) approach was employed as a solver for multivariable problems to optimize the predictive models. This comprehensive approach aimed to leverage the strengths of each algorithm and technique to develop robust and accurate predictive models for liquid flow rates through wellhead chokes.

In essence, the study's model development phase involved selecting appropriate algorithms, tuning parameters, and analyzing the performance of each model to ensure accurate predictions. By utilizing machine learning and optimization techniques, the researchers aimed to fill the gap in existing empirical correlations and provide a reliable method for predicting liquid flow rates in oil and gas production systems.

Result and Evaluation

The evaluation of predictive models for liquid flow rates through wellhead chokes involved statistical analysis and graphical tools to assess accuracy and reliability. Various statistical parameters such as average percent relative error (APRE), average absolute percent relative error (AAPRE), SD, root mean square error (RMSE), and coefficient of determination were calculated for comparison.

Machine learning models including AdaBoost-SVR, MARS, MLP and RBF were developed and evaluated against proposed correlations. The results showed that AdaBoost-SVR demonstrated the lowest AAPRE, indicating superior accuracy in predicting liquid flow rates. Graphical analyses, including cross-plots, cumulative frequency plots, trend plots, and error distribution plots, further confirmed the effectiveness of AdaBoost-SVR compared to other models and correlations.

Additionally, sensitivity analysis revealed the relative importance of input parameters, with choke size exerting the most significant influence on liquid flow rates. Outlier diagnostics using William's diagram indicated high model reliability, with the majority of data points falling within acceptable ranges. Furthermore, external validation of the AdaBoost-SVR model using an independent dataset demonstrated its robust predictive capability, even for fluid rates beyond the range used during modeling.

Overall, the AdaBoost-SVR model exhibited superior accuracy, reliability, and generalization ability compared to other models and correlations evaluated in the study. These findings underscored the effectiveness of machine learning approaches, particularly AdaBoost-SVR, in predicting liquid flow rates through wellhead chokes in oil and gas production systems.

Conclusion

In conclusion, the researchers significantly advanced the accuracy of predicting liquid flow rates through wellhead chokes in oil and gas production. By employing machine learning models and developing empirical correlations, the authors demonstrated superior performance compared to existing methods.

AdaBoost-SVR emerged as the most precise model, outperforming other intelligent models and correlations. The sensitivity analysis highlighted the importance of choke size in determining flow rates. Overall, the findings emphasized the critical role of accurate flow rate prediction in optimizing hydrocarbon recovery processes.