Hierarchical Phase-Adaptive Machine Learning Framework for Real-Time Non-Productive Time Prediction in Offshore Drilling

Abstract

NPT remains one of the largest financial burdens in the oil and gas industry, costing operators hundreds of millions of pounds per year and consuming 20–30% of total drilling time. The objective of this study was to develop a hierarchical phase-adaptive machine learning framework capable of dynamically adjusting feature importance and classification boundaries based on the type of drilling being conducted (vertical, curve, or lateral). As part of the validation process, the proposed method was applied to two wells located at the Volve Field in the Norwegian North Sea. Activity-level data were extracted from WITSML Daily Drilling Reports (DDRs) to construct a 31-feature transferable dataset using within-well normalized statistics, rates of change, and activity-based operational signals. Three phase-specific Random Forest classifiers were trained using SMOTETomek resampling on the primary well (15/9-F-10, 848 records, 90 NPT events). Cross-well validation on an entirely unseen well (15/9-F-9, drilled 18 months earlier) demonstrated that the phase-adaptive framework achieves F1=0.436 and AUC=0.798, representing improvements of +11.6 percentage points for F1 and +11.9 percentage points for AUC over static single-model baselines. Phase-specific analysis revealed that curve-phase operations are the most predictable (CV F1=0.966), while mud density and inclination variability emerge as universal NPT precursors across all drilling phases. The phase-adaptive framework achieved an average precision of 0.412, a 2.0-fold improvement over the static RF baseline, confirming the feasibility of transferable, operationally interpretable NPT prediction across wells and time periods.