Machine learning prediction of bottomhole flowing pressure as a time series in the volve field

Abstract

Bottomhole flowing pressure (BHFP) is a critical parameter in analyzing oil and gas well performance, production forecasting and reservoir management. This study is focused on obtaining feature combinations towards low-error prediction of time-series BHFP in two wells in the Volve field. Three machine learning (ML) models (support vector regression (SVR), a distance-based model; random forest (RF), a tree-based ensemble model and Long Short-Term Memory (LSTM), a type of recurrent neural network) are used for BHFP prediction in two wells of the Volve field. The data for each well was split such that the first 70% is used in training the model, the next 15% as validation data for selecting the optimal hyperparameters and the last 15% for testing the models. The train and validation sets were used to train the models before making predictions on the test sets. While the SVR and RF models reasonably predicted the BHFP in both wells with a maximum Mean Absolute Percentage Error (MAPE) of 5.0% and 4.3% respectively, the LSTM model performed best across both wells with the MAPE less than 2.9% in both wells. ML model performance was superior for the well with the data distributed more uniformly. The three feature combinations with superior ML model performance for BHFP prediction all have five features in common namely: bottomhole temperature, oil flow rate, gas flow rate, choke size, onstream hours. The workflow in this work can be adopted for fieldwide BHFP prediction.