Real-Time Rate of Penetration Prediction Analysis using LSTM Networks Under a Continuous Learning Scenario

Resumo

Recent technological advancements in data transmission and acquisition systems for drilling operations, along with the growing global energy demand, have led the Oil & Gas industry to pursue data-driven solutions in real-time to improve performance and reduce well construction costs. In this context, the Rate of Penetration (ROP) serves as a key drilling performance metric, reflecting the effective speed of the drill string as the bit penetrates the rock formation. Predictive models for ROP can be employed in real time to suggest appropriate operational parameters that optimize ROP, potentially reducing drilling time and overall operating costs. This work investigates the application of Long Short-Term Memory (LSTM) networks for ROP prediction within a continual learning scenario based on drilling data. LSTM networks are well-suited for this task due to their ability to model dependencies in sequential data. The results obtained using the LSTM model are compared with benchmark results from Multilayer Perceptron (MLP) networks. A scenario is simulated in which the models are incrementally trained with new data received during drilling, predicting ROP in subsequent intervals. Given the transient nature of drilling, abrupt changes in data distribution are expected and may significantly impact predictive performance. Therefore, the model's performance in the test intervals is analyzed in relation to statistical characterizations of the newly received data. The dataset used comprises public data from three wells in the Volve field in the North Sea. It includes both operational parameters and lithological information, the latter being particularly important for analyzing the results obtained in this study. Model performance is evaluated using the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The results demonstrate the potential of the adopted strategies using the LSTM model. It is expected that the comparative analyses presented in this study will contribute to key aspects of developing data-driven solutions for real-time drilling applications.