Data-Driven Prediction of Electrical Submersible Pump Run Life Using Leakage-Resistant Machine-Learning Validation

Abstract

Electrical submersible pumps (ESPs) are widely deployed to sustain production in mature and high-water-cut wells. Routine surveillance data capture equipment degradation and changing operating conditions and can therefore support remaining-useful-life (RUL) estimation and proactive maintenance planning. However, predictive performance may be substantially overstated when correlated daily observations from the same ESP run are distributed across training and test partitions. This study presents a comprehensive, leakage-resistant machine-learning (ML) and deep-learning (DL) framework that formulates ESP reliability prediction as two complementary tasks: continuous RUL regression and binary failure classification at 30-, 14-, 7-, and 1-day warning horizons. Sensor, production, equipment, and event-record data from Egypt's Western Desert were consolidated and temporally matched. Following data cleaning, preprocessing, and quality control, the final data set comprised 1,577 completed ESP runs—each defined by a matched installation and subsequent failure date—across 526 wells, together with 933,535 daily surveillance observations. Ten regression model families and six classification configurations were evaluated using a standardized pipeline incorporating engineering-based feature selection, training-only preprocessing, class-imbalance handling, hyperparameter optimization, cross-validation, early stopping, and learning-rate scheduling. On the held-out classification test set, the 30-day K-nearest-neighbor model achieved 98.98% accuracy, 0.889 failure-class precision, 0.909 recall, and an F1-score of 0.899. Random-row regression experiments yielded mean absolute errors of 3.61 days for K-nearest neighbors and 3.66 days for random forest; these results were treated as interpolation baselines rather than prospective deployment estimates. Blind field evaluation across five consecutive ESP runs from the fully withheld Well KX1 produced prediction errors of −27 to +1 days, corresponding to signed percentage errors from −9.75% to +0.24%, a mean absolute error of 12.4 days, a root mean square error of 15.0 days, and a mean absolute percentage error of 4.27%. Based on the operator-reported exposure of USD 250,000–300,000 per event, the five historical events represent USD 1.25–1.50 million in cumulative exposure. The pronounced class imbalance, predominance of nonfailure observations, and within-run correlation of daily records created substantial risks of information leakage and overly optimistic performance estimates. A multilayer validation framework was therefore implemented using failure-stratified, well-grouped, block-temporal, chronological, rolling-origin, and cluster-aware partitioning strategies, with all preprocessing fitted exclusively to the training partition. The deterioration observed under block-temporal and chronological evaluation confirmed that validation design materially influences apparent model performance. Overall, routine ESP surveillance data can support proactive diagnostic review, maintenance scheduling, ESP design optimization, and production planning, provided that performance claims are reported with the corresponding validation design, test population, and prediction uncertainty.