Intelligent Recognition of Kick Fluid Types Based on Time-Series Features: Method and Case

Summary

Deep and ultradeep drilling is often associated with frequent kick events and high operational risk. Inadequate control can readily trigger a range of downhole complications and may, in severe cases, lead directly to drilling failure. Rapid and accurate identification of kick fluid type is therefore critical for subsequent well-killing operations. At present, the identification of kick type still relies primarily on observations of return flow behavior or post-shut-in calculations based on wellhead parameters. Such approaches are time-lagged, may allow the risk to escalate, and increase the difficulty of subsequent well-killing treatment. In this study, early and accurate identification of kick type is formulated as a pattern-classification problem. Time-series features reflecting kick behavior are extracted from drilling-operation parameters, drilling-fluid parameters, and hydrocarbon-show parameters. Random forest (RF) is used to evaluate feature sensitivity (importance) and to construct feature vectors representing kick behavior, while the Savitzky-Golay filter is used for noise reduction. Common intelligent pattern-recognition methods are then analyzed in terms of their strengths, limitations, and applicability, and a long short-term memory (LSTM)-Transformer architecture is selected. The LSTM captures local features from multivariate long-sequence kick data, whereas the Transformer enhances global temporal dependency modeling before and after the kick. Accuracy (ACC), precision, recall, and F1-score are used to evaluate model performance, and a configuration with all four metrics above 90% is regarded as satisfactory. Key hyperparameters, including the number of LSTM layers, the number of hidden neurons, Transformer attention head dimension, the number of attention heads, optimizer type, and learning rate, are optimized via iterative grid search, resulting in an early and accurate intelligent recognition method for kick fluid types based on time-series features. Kick case data from the Tarim region were collected to establish a regional standard sample set of kick fluid types and their corresponding feature vectors. The optimized regional model uses two LSTM layers with 32 hidden neurons, a Transformer attention head dimension of 64, six attention heads, the Adam optimizer, and a learning rate of 0.001. The resulting model achieves an overall accuracy of 96.6%, a precision of 95.2%, and recall and F1-score values of 94.5% and 94.8%, respectively. Validation on case wells from the Junggar region yields a kick-type prediction accuracy of 93.9%, demonstrating good generalization and predictive capability. The proposed method can provide technical support for early on-site identification of kick fluid type and subsequent decision-making in accident handling.