Prediction of well performance parameters for wellhead choke flow optimization using MLP-GA

Accurate prediction of well fluid flow is essential for optimizing production; however, conventional multiphase flow models may be affected by measurement uncertainties and simplifying assumptions, which can limit their predictive reliability under complex operational conditions. In this study, a Multi-Layer Perceptron (MLP) integrated with a Genetic Algorithm (MLP-GA) was employed to predict key well performance parameters, including liquid production rate (QL). The dataset initially comprised 910 samples, of which 35 statistical outliers (≈3.85%) were identified and removed using robust statistical screening criteria to improve data consistency and reliability, resulting in 875 reliable observations. Four critical input parameters were considered: wellhead pressure (Pwh), choke size (D64), basic sediment and water content (BS&W), and gas-to-liquid ratio (GLR). A comprehensive evaluation of data partitioning schemes was performed, varying the test set from 10% to 90% in 10% increments, to identify the optimal training-to-test ratio. The optimal configuration consisted of 80% of the data allocated to model development, including training (70%) and validation (10%), and the remaining 20% reserved for independent testing. This split provided the best performance, achieving R2 values of 0.9983, 0.9989, and 0.9914 for training, validation, and testing datasets, respectively, with mean absolute errors of 0.1603, 0.1910, and 0.3333. The optimized MLP-GA model employed two hidden layers (73 and 18 neurons). The number of neurons in the hidden layers was automatically determined by the Genetic Algorithm (GA) through a systematic search within predefined bounds, aiming to minimize the validation error and ensure stable convergence. A tanh activation function, an adaptive learning rate schedule, and a learning rate of 0.0610. Further interpretability analyses using SHAP, ICE, PDP, and LIME were conducted to identify the most influential features and understand their impact on individual and global model predictions. The results demonstrate that the proposed MLP-GA framework provides highly accurate and interpretable predictions of well performance parameters, offering a practical tool for wellhead choke flow optimization and informed decision-making in production operations.