Automated Well Status Identification in US Offshore Operations Using Machine Learning

Well status identification is crucial for promoting safety, optimizing efficiency, and helping protect the environment in offshore operations. Additionally, it provides valuable input for operations, metering, allocations, surveillance, engineering, etc. This process involves interpreting data from numerous sensors installed on the well and production platform, which is traditionally labor-intensive. This paper presents a machine learning–based approach to automating well status classification on a US offshore asset, leveraging high-resolution time-series data to improve accuracy, consistency, and scalability. Automating well status identification can be challenging, even on wells with multiphase flowmeters, which require periodic calibration. Additionally, not every well has a multiphase flowmeter. Therefore, a more precise approach is necessary to accurately determine the shut-in period and the associated metrics. This required developing a machine learning model for binary classification of well status (flowing or shut in) using high-frequency well-level time-series data. After the machine learning models were developed, a sliding window approach was applied to transform high-frequency sensor data into structured samples, effectively treating each window as a time-series "image." Each window was labeled based on its final data point to enable early detection of well shutdowns. To address the imbalance caused by the predominance of flowing data, data augmentation techniques were employed. Multiple architectures of One-Dimensional Convolutional Neural Network (1D CNN) and Extreme Gradient Boosting (XGBoost) models were tested, and hyperparameter tuning—including optimization of the sliding window size—was performed to enhance model accuracy. Initially, individual models were trained for each well; later, a unified model was developed using data from all 11 wells to support labeling in new wells with limited historical data. The best-performing model achieved average accuracies of 99.92% during training, 99.84% for validation and 99.86 for blind test. Additionally, the classification framework was integrated with a signal processing–based peak detection algorithm to accurately identify ramp-up and ramp-down periods. This enabled further capabilities such as automated productivity index calculation, and detection of anomalous buildups and ramp-up behaviors. The integration of machine learning models with peak detection algorithms for real time sensor data analysis to automatically determine well status and extract necessary metrics makes this study novel and highly relevant for offshore applications. It also demonstrates how Artificial Intelligence (AI) can automate tasks, reduce manpower, and enhance efficiency and safety in offshore operations.