Deep Learning Method for Improving Rate of Penetration Prediction in Drilling

Abstract

The urgent global need to reduce CO2 emissions necessitates the development of sustainable power generation sources. Geothermal power emerges as a renewable and dependable energy option, harnessing the Earth’s natural heat sources for electricity generation. Unlike other renewables, geothermal energy offers uninterrupted power, immune to weather conditions. However, its efficiency hinges on technological innovation, particularly in the challenging realm of geothermal drilling. Rate of penetration (ROP) is a crucial drilling performance metric, and this study explores how deep learning models, particularly transformers, can optimize ROP prediction. Leveraging data from Utah Frontier Observatory for Research in Geothermal Energy (FORGE), we analyze the relationship between drilling parameters and ROP. Traditional drilling optimization methods face limitations, as drilling dysfunctions can disrupt the linear relationship between ROP and weight on bit (WOB). We propose a dynamic approach that allows adapting drilling parameters in real time to optimize ROP. Our experiments investigate optimal sampling intervals and forecast horizons for ROP prediction. We find that a 60-second sampling interval maximizes the transformer model’s forecasting accuracy. Additionally, we explore retraining to fine-tune models for specific wells, improving forecasting performance. Our transformer-based ROP forecaster outperforms deep learning models, achieving a low overall 5.22% symmetrical mean average percentage error (SMAPE) over a forecast horizon of 10 minutes. This model offers opportunities for cost-effective drilling optimization, with real-time accuracy, speed, and scalability. Future work will focus on larger data sets and integration with drilling automation systems to further enhance the model’s practicality and effectiveness in the field.