An interpretable model for real-time prediction of bottomhole flowing pressure in deviated wells: a hybrid GA-ANN approach

Flowing bottomhole pressure (FBHP) is an important parameter in forecasting well productivity. Machine learning algorithms have been implemented to predict FBHP. However, most of the models are black-box models without explicit mathematical equations. These models focused on precision but not interpretability and transparency. In this study, 354 datasets were collected from five deviated wells from a field located in the Gulf of Guinea to develop models for estimating FBHP. The dataset consists of easily measurable parameters obtained during production including oil flow rate, gas flow rate, water flow rate, gas oil ratio, water oil ratio, tubing head pressure, and FBHP. K-Nearest Neighborhood (KNN) and Artificial Neural Network (ANN) were employed to train the model. A hybrid model was developed by optimizing the weights and biases extracted from the trained ANN using Genetic algorithm (GA-ANN) with RMSE as the fitness function. GA-ANN was found to outperform the KNN and ANN models with MSE, RMSE, AAPRE, and R2 of 0.0115, 0.1074, 0.1177, 0.9883, respectively. The mathematical equation of the GA-ANN model was then extracted using white-box technique. To validate the proposed model, the dependence of the correlation on its input variables were examined through trend analysis. Uncertainty analysis was carried out via the Monte Carlo Dropout-based uncertainty quantification and the results revealed that aleatoric and epistemic uncertainties were 96.3% and 3.7%, respectively showing the model’s reliability. To the best our knowledge this is the first time an interpretable intelligent model is proposed for predicting FBHP in deviated wells.