Impact of Automated Time Series Feature Engineering on Enhancing Incident Prediction Rates in Geothermal Drilling

Abstract

Improving overall drilling efficiency in geothermal systems involves two key strategies: increasing the rate of penetration (ROP) and minimizing non-productive time (NPT). Drilling operations generate vast amounts of time-series data, including parameters like weight on bit, rotation speed, mud flow rate, and temperature. This data can be highly variable and noisy, making it challenging to interpret without sophisticated analysis. Machine learning (ML) offers significant promise in providing a data-based decision support system for drilling operations in addressing geothermal challenges through its predictive modelling and adaptive learning capabilities. By applying machine learning algorithms to both the raw data and engineered features, the system can identify patterns that precede hole-related non-productive time (NPT) incidents, such as stuck pipe situations or borehole instability. This paper compares the improvement in incident prediction rate of the model arising from the use of automated time-series feature engineering techniques as compared to the traditional naïve, manual, and domain-specific feature generation practices. Early detection is crucial in taking corrective actions before problems escalate. By addressing issues before they result in significant NPT, the system can lead to more efficient drilling schedules and less wasted time, ultimately contributing to improved operational performance. As this system matures, it has the potential for widespread deployment across various drilling operations.