Data-driven soft sensing model for production system of offshore gas fields based on deep learning

In view of the problems of great difficulty, high cost and low reliability in obtaining the operation data on the key state variables of production system in offshore natural gas fields, research on soft sensing was carried out, and a dynamic data-driven estimation model of flow and pressure was established, providing the central controllers an online monitoring tool for safety analysis of the system. Combining the dynamic and static samples, a Dense Neural Network model library of Nonlinear Auto-Regressive with Exogenous Inputs(DNN-NARX model), comprising the black box and grey box, was established with the black-box identification technique based on deep learning, as well as the parameter correction technique based on transfer learning, approximately describing the dynamic flowing characteristics of gas production well, so as to estimate the single well flow and wellhead pressure. The simulation results of the dynamic DNN-NARX black-box and grey-box model, the DNN-NARX model and the traditional Multiple-Layer Perception-NARX model(MLP-NARX model)were compared by examples of calculation. The results indicate that both of the accuracy and calculation time of DNNNARX model and MLP-NARX model satisfy the requirements of online estimation. Therein, the dynamic DNN-NARX grey-box model shows the remarkable advantages of higher immunity from interference and generalization capability. Thus,the proposed model has high engineering applicability, providing good referential significance to the soft sensing problems in the field of offshore oil and gas production.