Modelling the Oil-in-Water Separation Dynamics in a De-Oiling Hydrocyclone System Using LSTM Neural Network

By deploying the online Oil-in-Water (OiW) sensors in a de-oiling hydrocyclone system used for produced water treatment processes in offshore oil & gas production, this work investigated modelling of the complicated separation dynamics inside the hydrocylone system using the Long-Short-Term Memory Neural Network (LSTM-NN). The purpose of this modelling is to predict the hydrocyclone's transient de-oiling efficiency in a high level of accuracy. Thereby the hydrocyclone system can be optimally controlled subject to different operating conditions. The acquisition and analysis of the data obtained from a lab-scaled pilot plant is introduced. Two types of LSTM-NN configurations are proposed, and the hyper-parameter tuning as well as training and validation results, are discussed in details. The results exhibit that the relative concentration of OiW, which correlated with the de-oiling efficiency, can be predicted in a quite accurate level using two types of measurements, i.e., the opening degrees of cyclone's underflow and overflow control valves, both the hydrocyclone's inlet/water-outlet OiW concentration measurements. The best model can achieve a normalized RMSE 83,62% accuracy in the validation test. One of our next step is to cooperate the LSTM-NN model into the model predictive control framework to design some optimal control solution for de-oiling hydrocyclone systems.