Downhole Temperature Estimation in Geothermal Wells Using a Deep Learning Model Based on LSTM Neural Networks

Abstract

Abstract Accurate real-time modeling of downhole temperature (DHT) in high-temperature geothermal wells is crucial for proactive temperature management and preventing thermally induced drilling issues. While existing numerical models excel at precise DHT prediction, their complexity and long computational times render them unsuitable for real-time operations. This study introduces an innovative machine-learning model that overcomes these limitations, enabling real-time DHT monitoring in geothermal wells. To create a robust dataset simulating DHT behavior under various drilling conditions in FORGE wells, we leveraged a sophisticated thermo-hydraulic model validated with Utah FORGE field data. This dataset, comprising thousands of data points, served as the training ground for a Deep Long Short-Term Memory (DLSTM) model. The DLSTM model, designed to capture the intricate non-linear relationship between DHT and drilling parameters, was fine-tuned using a Bayesian algorithm that efficiently optimized model settings based on past evaluation results. In a comprehensive evaluation, the model was rigorously tested against previously unseen scenarios to assess its strengths and limitations. Additionally, a parametric analysis was conducted to validate the model's predictions against established temperature management techniques reported in the literature. Key findings indicate that variables such as DHT from previous time steps, mud type, and wellbore horizontal lateral length play pivotal roles in DHT estimation. The developed DLSTM model exhibits exceptional precision, stability, and generalizability in predicting DHT during both circulation and pump-off (no circulation) scenarios. Its architecture, featuring two LSTM layers with numerous cells, augmented by an additional hidden layer of artificial neural networks, effectively mitigates overfitting issues commonly encountered with conventional neural networks when dealing with extensive time-series drilling data. Across all scenarios, the model utilizes a 3-minute lag time of time series data to accurately simulate DHT in geothermal wells, achieving a Mean Absolute Error (MAE) consistently below 1°C in most cases. Notably, the DLSTM model successfully captures the dynamic behavior of bottomhole circulating temperature (BHCT) and DHT build-up during pump shut-off periods in various drilling scenarios. Furthermore, the model's predictions regarding the impact of various cooling strategies on downhole temperature align well with the established understanding of temperature management techniques in geothermal wells. Our developed machine learning model offers a reliable and automated solution for DHT prediction and real-time heat management in geothermal wells. It serves as a practical and significantly faster alternative to complex, time-consuming physics-based models. This research underscores the DLSTM's potential for capturing intricate time-dependent relationships between input features (drilling parameters) and outputs, opening avenues for its application in diverse drilling contexts beyond geothermal wells.