Artificial Intelligence for the Online Prediction of the Cool-down Time in a Subsea Pipeline After an Unplanned Shutdown

Hydrates formation in subsea pipelines is one of the main reliability concerns for flow assurance engineers. A fast and reliable assessment of the Cool-Down Time (CDT), the period between a shut-down event and possible hydrates formation in the asset, is of key importance for the safety of operations. Existing methods for the CDT prediction are highly dependent on the use of very complex physics-based models that demand large computational time, which hinders their usage in an online environment. Therefore, this work presents a novel methodology for the development of surrogate models that predict, in a fast and accurate way, the CDT in subsea pipelines after unplanned shutdowns. The proposed methodology is, innovatively, tailored on the basis of reliability perspective, by treating the CDT as a risk index, where a critic CDT threshold (i.e. the minimum time needed by the operator to preserve the line from hydrates formation) is considered to distinguish the simulation outputs into high-risk and low-risk domains. The methodology relies on the development of a hybrid Machine Learning (ML) based model using datasets generated through complex physics-based model’ simulations. The hybrid ML-based model consists of a Support Vector Machine (SVM) classifier that assigns a risk level (high or low) to the measured operating condition of the asset, and two Artificial Neural Networks (ANNs) for predicting the CDT at the high-risk (low CDT) or the low-risk (high CDT) operating conditions previously assigned by the classifier. The effectiveness of the proposed methodology is validated by its application to a case study involving a pipeline in an offshore western African asset, modelled by a transient physics-based commercial software. The results show outperformance of the capabilities of the proposed hybrid ML-based model (i.e., SVM + 2 ANNs) compared to the classical approach (i.e. modelling the entire system with one global ANN) in terms of enhancing the prediction of the CDT during the high-risk conditions of the asset. This behaviour is confirmed applying the novel methodology to training datasets of different size. In fact, the high-risk Normalized Root Mean Square Error (NRMSE) is reduced on average of 15% compared to the NRMSE of a global ANN model. Moreover, it’s shown that high-risk CDT are better predicted by the hybrid model even if the critic CDT, which divides the simulation outputs in high-risk and low-risk values (i.e. the minimum time needed by the operator to preserve the line from hydrates formation), changes. The enhancement, in this case, is on average of 14.6%. Eventually, results show how the novel methodology cuts down by more than one hundred seventy-eight times the computational times for online CDT predictions compared to the physics-based model.