Abstract:

In order to further enhance the early abnormal warning capability of natural gas compressor units, a novel method was proposed based on the VMD (Variational Mode Decomposition) algorithm, Informer algorithm, 3σ criterion, and Correlation-Weight optimization. A predictive model was constructed using the Informer architecture, in which the monitoring data were decomposed by VMD algorithm into multi-scale features of different frequencies to serve as model inputs. During the training process, the weight coefficients between each decomposed component and the original signal were calculated to optimize and adjust the internal model parameters. Furthermore, the prediction reconstruction results were combined with the 3σ statistical criterion to further improve the warning performance. Two segments of normal and abnormal pressure differential monitoring data from field compressor units were collected for experimental validation. The results show that, compared with other prediction methods, the proposed warning method achieves the lowest prediction errors. For the prediction of normal box pressure differentials, the errors are reduced by 66.67%-71.43% (Mean Squared Error, MSE), 36.67%-45.45% (Mean Absolute Error, MAE), 40.17%-45.42% (Root Mean Squared Error, RMSE), and 36.57%-45.72% (Mean Absolute Percentage Error, MAPE). For the abnormal inlet pressure differentials, the errors are reduced by 64.43%-71.12% (MSE), 44.02%-52.27% (MAE), 40.36%-45.53% (RMSE), and 37.24%-47.79% (MAPE). The proposed method exhibits superior prediction accuracy in both detailed and trend features. In the test set, it provides anomaly warning 60 minutes in advance, thereby improving the reliability of safe and stable operation of the compressor units.

Key words: variational correlation-weight Informer(VCW-Informer), natural gas compressor, monitoring data, abnormal early warning, deep learning, signal decomposition