Session: 04-04-01 Flow Assurance & Subsea Equipment

Paper Number: 79494

79494 - Time Series Prediction of the Trend of Hydrate Risk Using Principal Component Analysis and Deep Learning 

Hydrate risk strategy is critical concern in offshore gas and oil production system. Several empirical models were employed to predict the hydrate formation behaviors related plugging risk. However, this empirical approaches have limitation in becoming universally used due to its dependency on geometries and fluid characteristics. In general, the probability of hydrate plugging in the pipeline could increase when the hydrate particle distribution changed from homogeneous to heterogeneous. However, hydrate kinetic behaviors are statistical and nonlinear relationship on the dependent variables, which means it is difficult to develop the model to describe it behavior.

In this work, time series prediction using data-driven methods are applied rather than these model-based methods to analyze the kinetic experiment data during the hydrate formation. Classical time series models use the statistical methods which is not able to be applied on the non-stationary data or data without patterns. On the other hand, deep learning method can provide the prediction of non-stationary or abrupt events.

Several deep learning models like LSTM(Long Short-Term Memory), GRU(Gated Recurrent Unit), ARLSTM(Auto Regressive Long Short-Term Memory) were used to be trained based on the lab-scale experiment data to make the real time prediction. The transition trend of hydrate formation from homogenous to heterogenous particles was predicted by using the model. Prediction was made on hydrate risk indicator, which is PCA(Principal Component Analysis) treated sensor data including pressure, temperature, relative torque and others. Two group of data with different mixing rate was used in prediction to examine the universal applicability of the model.

Training with dataset under similar experimental condition showed better result than training with dataset under different conditions, but its affects less if dataset number gets larger. The results suggested that the deep learning techniques incorporating with time series prediction could be the promising method for hydrate risk management.

Presenting Author: Nayoung Lee Seoul National University

Authors:

Nayoung Lee Seoul National University
Hyunho Kim National University of Singapore
Yutaek Seo Seoul National University