Machine Learning Models to Predict Pressure at a Coiled Tubing Nozzle's Outlet During Nitrogen Lifting

Abstract

Abstract Optimizing bottom hole pressure is crucial for successful nitrogen lifting during clean-out phases after well interventions. Precisely predicting bottom hole pressure is vital for evaluating Inflow performance relationship (IPR) and optimizing operational parameters (e.g., nitrogen injection rate, Run in hole (RIH) speed, and CT depth) in real-time. Multiphase flow around the CT complicates physics-based pressure estimation. This effort aims to develop accurate machine learning models for predicting bottom-hole pressure at the CT nozzle during nitrogen lifting, especially in wells lacking down-hole gauges. A machine learning model is developed using readily available parameters typically gathered during nitrogen lifting operations, which include wellhead flowing pressure, wellhead flowing temperature, bottom hole temperature, oil density, water salinity, production rate, water cut percentage, gas-oil ratio, nitrogen rate, gas gravity, and CT depth as inputs. This model is trained utilizing measured bottom-hole pressure data acquired from deployed memory gauges, serving as the model's outputs. Nine distinct machine learning algorithms—Gradient Boosting, AdaBoost, Random Forest, Support Vector Machines (SVMs), Decision Trees, K-Nearest Neighbor (KNN), Linear Regression, Neural Network, and Stochastic Gradient Descent (SGD)—are meticulously developed and fine-tuned utilizing data streams obtained from diverse well operations across 235 wells through data acquisition systems. This dataset is split into two subsets: 80% for training the algorithms and 20% for rigorously testing their predictive capabilities. Two cross-validation processes (K-fold and random sampling) are used to assess the performance of machine learning models. The outcomes of the top-performing machine learning models, specifically Gradient Boosting, AdaBoost, Random Forest, SVMs, and Decision Trees, reveal remarkably low mean absolute percent error (MAPE) values when comparing their predictions of coiled tubing (CT) nozzle outlet pressure to actual measurements. These MAPE values stand at 2.1%, 2.7%, 2.8%, 6.6%, and 5%, respectively. Additionally, the correlation coefficients (R2) for these models are notably high, with values of 0.936, 0.906, 0.896, 0.813, and 0.791, respectively. Furthermore, machine learning models offer distinct advantages over conventional vertical lift performance curve correlations, as they do not necessitate routine calibration. Beyond this, these models demonstrated their ability to predict bottom-hole pressure across various operations using data that the models had never encountered during training. Predictions were compared to actual measurements, showing a strong alignment between the model's predictions and real-world bottom-hole pressure data. This paper introduces novel insights by demonstrating how using a machine learning model for predicting CT nozzle outlet bottomhole pressure across diverse pumping conditions can enhance ongoing nitrogen lifting operations. Utilizing machine learning models offers a more efficient, rapid, real-time, and cost-effective alternative to calibrated vertical lift performance correlations. Furthermore, these models excel in accommodating a wide spectrum of reservoir fluid characteristics.