Abstract：
During the shutdown of a multi-product pipeline, the pipeline pressure will drop due to the temperature difference
inside and outside the pipeline. In addition, some abnormal accidents, such as oil theft will also reduce the pipeline pressure.
Therefore, it is difficult to distinguish whether there has been an oil theft accident or not. When the pipeline pressure drops, on
site personnel often mistakenly think that abnormal accidents have happened, such as pipeline leakage or oil theft, increasing the
management burden on the site. In order to achieve the goal of real-time monitoring of pipeline pressure changes and effective
guidance of on-site management, we have carried out research on shutdown pressure prediction of a multi-product pipeline.
First, based on the mechanism model, the influencing factors of the pipeline pressure change (shutdown time, oil temperature,
and ambient temperature) were determined by an empirical formula. Based on pipeline SCADA data and weather condition
data, a sample database of shutdown pressure was constructed. In order to improve the prediction accuracy, the characteristics
of the time series of pressure change were considered, and a particle swarm optimization (PSO) algorithm is used to optimize
the hyperparameters of a long short-term memory (LSTM) model. Finally, a pressure prediction model is established for a
multi-product pipeline during the shutdown. Taking mean absolute error (MAE), root mean squared error (RMSE), and mean
absolute percentage error (MAPE) as the model indicators, three domestic multi-product pipelines with different shutdown peri
ods were taken as examples and compared with other prediction models such as basic LSTM, support vector regression (SVR),
decision tree (DT), random forest (RF), and artificial neural network (ANN). The results of the examples show that the pressure
prediction model based on PSO-LSTM has the best effect, especially when the duration of shutdown is shorter, and its effect
is more prominent. For pipeline A, RMSE, MAE, and MAPE are 0.009, 0.008, and 0.167 respectively. For pipeline B, RMSE,
MAE, and MAPE are 0.008, 0.007, and 0.309 respectively. For pipeline C, RMSE, MAE, and MAPE are 0.018, 0.015, and 0.128
respectively. The pressure prediction model established in this paper can dynamically predict pipeline pressure changes, realize
real-time monitoring of pipeline pressure, and improve the efficiency of on-site operation management. When the difference
between the predicted value and the detected one is large, it can be considered that an abnormal condition has occurred in the
pipeline and needs to be checked on site.
