Data-Driven Modelling of Oil Flow Rates in Vertical Wells: Performance Comparisons of Neural Network and Multigene Genetic Programming Approaches

Abstract

Accurate prediction of oil production rates remains challenging, especially in fields where direct measurement is not feasible. To address this issue, data-driven models were developed using neural networks (NN) and multigene genetic programming (MGGP) to provide real-time estimations of oil flow rates in vertical wells using routinely measured field parameters. The models were trained and tested on 70% and 30% of a dataset comprising 1,893 entries, respectively. The input variables for the models include oil API gravity, downstream temperature, upstream temperature, upstream pressure, and choke size. These developed models are simplified, reproducible, and suitable for practical application. Model performance was quantitatively evaluated using the correlation coefficient (R), the coefficient of determination (R2), the mean squared error (MSE), and the root mean square error (RMSE). The NN model demonstrated superior predictive capability, achieving R = 0.990, R2 = 0.980, MSE = 0.0001, and RMSE = 0.011, while the MGGP model achieved R = 0.980, R2 = 0.960, MSE = 0.0004, and RMSE = 0.020. Validation using an independent dataset further confirmed the robustness of the NN model, which outperformed the MGGP model across all evaluation metrics. Additional contributions of this study include sensitivity analysis of input variables, derivation of explicit predictive correlations, and evaluation of computational efficiency. Both models exhibited low computational cost, supporting real-time and field-scale applicability. Thus, the NN model outperforms the MGGP model in estimating oil field flow rates. Therefore, production engineers are recommended to utilise an NN-based model as a decision-support tool for estimating oil flow in vertically flowing wells.