Machine Learning-Based Production Rate Forecasting from ESP Parameters with SHAP Explainable AI

Abstract

Abstract Electrical submersible pump (ESP) wells often operate without continuous flow metering, forcing engineers to infer rate from sparse tests and noisy telemetry. This study builds a virtual flow metering workflow that predicts daily oil rate directly from ESP operating parameters and explains each prediction using SHAP (SHapley Additive exPlanations). A field dataset of 4,700 daily records was used with eight inputs spanning downhole pressures and temperatures, surface electrical measurements, fluid API gravity, and submergence. Three tree-ensemble regressors (ExtraTrees, XGBoost, and LightGBM) were trained and benchmarked using an 80/20 train-test split and 5-fold cross-validation. All models achieved strong generalization; the best model (ExtraTrees) reached R2 = 0.977 on the test set with AAPE = 1.55% (MSE = 5.4×103). Cross plots cluster tightly about the 45° line, and residuals are centered near zero, indicating minimal bias. SHAP analysis shows that intake temperature (Ti) and VSD output current are the dominant drivers of predicted rate: high Ti systematically reduces predicted production, while stronger electrical drive conditions increase it. The explanations are consistent with ESP physics and provide actionable diagnostics for operations. The results demonstrate that accurate, transparent rate forecasting can be achieved from routine ESP data, enabling practical real-time production surveillance and optimization.