Reducing NPT Using a Novel Approach to Real-Time Drilling Data Analysis

Abstract

Abstract Early detection and characterization of anomalous events during drilling operations are critical to avoid costly downtime and prevent hazardous events, such as a stuck pipe or a well control event. A key aspect of real-time drilling data analysis is the capability to make precise predictions of specific drilling parameters based on past time series information. The ideal models should be able to deal with multivariate time series and perform multi-step predictions. The recurrent neural network with a long short-term memory (LSTM) architecture is capable of the task, however, given that drilling is a long process with high data sampling frequency, LSTMs may face challenges with ultra-long-term memory. The transformer-based deep learning model has demonstrated its superior ability in natural language processing and time series analysis. The self-attention mechanism enables it to capture extremely long-term memory. In this paper, transformer-based deep learning models have been developed and applied to real-time drilling data prediction. It comprises an encoder and decoder module, along with a multi-head attention module. The model takes in multivariate real-time drilling data as input and predicts a univariate parameter in advance for multiple time steps. The proposed model is applied to the Volve field data to predict real-time drilling parameters such as mud pit volume, surface torque, and standpipe pressure. The predicted results are observed and evaluated. The predictions of the proposed models are in good agreement with the ground truth data. Four Transformer-based predictive models demonstrate their applicability to forecast real-time drilling data of different lengths. Transformer models utilizing non-stationary attention exhibit superior prediction accuracy in the context of drilling data prediction. This study provides guidance on how to implement and apply transformer-based deep learning models applied to drilling data analysis tasks, with a specific focus on anomaly detection. When trained on dysfunction-free datasets, the proposed model can predict real-time drilling data with high precision, whereas when a downhole anomaly starts to build, the significant error in the prediction can be used as an alarm indicator. The model can consider extremely long-term memory and serve as the alternative algorithm to LSTM. Furthermore, this model can be extended to a wide range of sequence data prediction problems in the petroleum engineering discipline.