Data-Driven Estimation of Flowing Bottom-Hole Pressure in Petroleum Wells Using Long Short-Term Memory

Monitoring bottom-hole pressure (BHP) during flowing periods in petroleum wells is critically important for reservoir management, formation evaluation, lift optimization, and flow assurance. The advent of Permanent Downhole Gauges (PDGs) has changed paradigms regarding the availability and accessibility of this information. However, the costs associated with installing or replacing these gauges can sometimes be prohibitive. In this work, we propose a data-driven framework to fill the information gap when PDG data is unavailable. We used surface measurement data and wellhead gauge readings as inputs and train machine learning methods to accomplish this task. We implemented and evaluated several configurations of Long Short-Term Memory (LSTM) and compared their results with non-deep learning methods as Multi-Layer Perceptron (MLP) Neural Networks and Ridge Regression, taking as case study a real dataset from an offshore oilfield in the Brazilian Pre-salt. We show that our framework can provide reliable estimated values of flowing BHP, with error metrics MAPE and SMAPE consistently under 3%. We highlight the following main contributions: (1) the utilization of time-related data; (2) the development of a framework applicable across a wide range of reservoir and flow conditions; and (3) the potential for practical applications, including real-time monitoring.