A hierarchical early kick detection method using a cascaded GRU network

Abstract

The lagged response of some monitoring parameters leads to alarm delays in a multi-parameter kick warning system. In order to improve the timeliness of kick warning, a hierarchical early kick detection method using a cascaded gated recurrent unit (GRU) network is proposed. In this method, the GRU is used as the basic unit to detect abnormal parameter changes, a graded kick warning model is constructed, in which a risk assessment of low, moderate, high likelihood is took depending on the number of parameters that have been abnormal at different moments. The cascaded network monitors parameters of different sensitivities in groups, and obtaining better data fitting effect by reducing the number of parameters detected by each level of the network, thus achieving a more accurate classification of the kick likelihood. In addition, a new parameter priority indicator, that measures the change rate and reaction time of the parameter using the variation of the parameter, is proposed for parameter responsiveness ranking. The test results using the field data from 22 wells show that the proposed method achieves the correct grading of low risk to high risk. Compared with the conventional GRU model, the kick recognition-accuracy is improved by 5.88%, while the alarm time is 1.5 min earlier on average.