An intelligent data-driven model for virtual flow meters in oil and gas development

Abstract

In this work, Back Propagation (BP) neural network, Long Short-Term Memory (LSTM) network and Random Forest algorithm are applied to establish an intelligent data-driven model for virtual flow meters in oil and gas development. The actual data of two oil wells in an offshore oil field in the South China Sea are used to construct a dataset. Feature engineering and parameter optimization are carried out in sequence, and three data-driven models are established. The model utility is evaluated in terms of model stability and data volume requirement. Among the three models, the LSTM model shows the highest accuracy with a MAE (Mean Absolute Error) of 3.9 %, the highest stability, and moderate data volume requirement. The BP network exhibits the lowest accuracy with a MAE of 12.1 %, the lowest stability, and the smallest data volume requirement. Random forest exhibits moderate accuracy, high model stability and highest data volume requirement. Finally, the transfer learning model based on LSTM model and BP network is proposed and tested. The results show that the data volume requirement of the transfer learning model are reduced by half when the accuracy of the prediction results is not much different from that of the LSTM model. The proposed transfer learning model can use available data more efficiently and improve the generalization ability of the model. This work provides further insights of data-driven model application in virtual flow meters, which is of great significance for the intelligent green oil and gas engineering.