Multi-Output Machine Learning for Simultaneous Prediction of Oil, Gas, and Water Rates: Effect of Well Data Partitioning on Model Performance

Abstract Machine learning models for petroleum production forecasting are routinely benchmarked using random data splitting, yet production records are temporally ordered and well-clustered, making this practice susceptible to data leakage. A multi-output machine learning workflow for simultaneous prediction of oil, gas, and water production rates is developed. The proposed multi­machine learning workflow systematically quantifies the impact of evaluation protocols on model performance. Five algorithms (Linear Regression, Random Forest, Extra Trees, XGBoost, and HistGradientBoosting) were trained and evaluated under three partitioning protocols (random, chronological, and well grouped splitting), yielding fifteen model–protocol combinations on the SPE DSEATS 2024 Africa Region dataset (5,945 observations; 5 wells). Under random splitting, all three ensemble models exceeded R2 = 0.93. Under the deployment-credible chronological protocol, the best model (HistGradientBoosting) achieved R2 = 0.5638, representing a decline of 0.37 R2 points, while all models collapsed under well-grouped splitting. The strongest model spans 1.23 R2 points across protocols, compared with a maximum 0.22-point gap across models within any single protocol. SHapley Additive exPlanations (SHAP) TreeExplainer was applied to the best chronological model to produce per-phase feature attributions. A Streamlit dashboard delivers batch inference, per-target SHAP attribution, and actual-versus-predicted trend visualization. The results show that evaluation protocol choice has a five- to six-fold larger influence on reported accuracy than model selection on this temporally structured dataset.