Development of Machine Learning for Lost Production Opportunity Prediction and Production Optimization Through Workover Scheduling in Pertamina Hulu Sanga Sanga Field, Sanga-Sanga Block of East Kalimantan, Indonesia

Abstract

Lost Production Opportunity (LPO) measures production that cannot be realized because a well or facility is constrained. In many field-planning workflows, LPO and workover schedules are still estimated from historical averages and a fixed production uplift for each job. Although practical, these assumptions do not represent the nonlinear interaction among production behavior, equipment failures, downtime, and intervention timing. This work examines whether those interactions can be learned from field data and used to improve the monthly scheduling of rigless workover jobs.Historical production, Planned Downtime (PDT), Unplanned Downtime (UPDT), and operational event records from Field ‘X’ were compiled at monthly resolution. XGBoost regression models were trained using a chronological data split and evaluated with MAE, RMSE, sMAPE, and actual-versus-predicted plots. The resulting LPO forecasts and feature-importance rankings were then incorporated into a Decision Support System (DSS). Together with workover capacity, expected uplift, job cost, commodity price, and operational constraints, the DSS assigns a monthly risk level and recommends whether workover activity should be increased, selected cautiously, or reduced.The gas and oil models produced sMAPE values of 16.79% and 19.06%, respectively. In the 2026 simulation, DSS-based scheduling increased gas production from 83.43 to 84.56 MMscfd and oil production from 9,650 to 9,805 BOPD relative to the conventional schedule. These results suggest that field data can be used not only to estimate LPO, but also to place rigless workover jobs in months where their production benefit is less exposed to operational disruption.