Forecasting the risk of formation damage (colmatation) with a 14-day horizon using calibrated probabilistic models

Abstract

Formation damage (colmatation) is a major contributor to production losses and increased operating expenditures, yet routine diagnostics are often reactive and provide limited lead time for low-cost corrective actions. This study develops a reproducible 14-day early-warning workflow that forecasts well-calibrated event probabilities (not only risk rankings) and links them to an economically justified operating threshold. The target event is defined as a productivity-index decline below an engineering limit, IC = Qo_fact / Qo_theor < 0.90, occurring within a 14-day horizon. Daily operational time series (water cut, net GOR, wellhead pressure, choke setting, GLR, and production rates) are transformed into a leakage-safe dynamic feature set using lags, rolling statistics, and trend descriptors computed strictly from information available at the forecast time. Model development follows rolling-origin (time-aware) backtesting with out-of-fold probability calibration (Platt scaling and isotonic regression). Performance is assessed in terms of discrimination (PR-AUC/ROC-AUC) and calibration (Brier score, expected calibration error, and calibration plots), while the operating threshold is chosen by minimizing expected decision cost under asymmetric error penalties and validated via decision-curve and lift analyses. The calibrated models provide operationally meaningful probabilities that align with empirical event rates across temporal windows, and the selected threshold yields positive net benefit compared with treat-all/treat-none baselines. The proposed pipeline is suitable for integration into a decision-support workflow with explicit recommendations for temporal validation, calibration, cost-based thresholding, and drift monitoring.