| Title | Field Validation of a Machine Learning Model For Real-Time Prediction of Geothermal Well Output using Differential Pressure Measurements |
|---|---|
| Authors | Agata Rostran Largaespada, Egill Júliusson, Hilmar Már Einarsson, María Sigríður Guðjónsdóttir |
| Year | 2026 |
| Conference | World Geothermal Congress |
| Keywords | |
| Abstract | Accurately measuring the output of geothermal wells producing two-phase flow remains a major challenge in the geothermal industry. Traditional methods, such as tracer dilution or separator tests, are costly, intrusive, and infrequently performed. In previous work, a machine learning model was developed to predict steam quality from three-tap differential pressure (DP) measurements over an orifice plate. Steam quality estimates were then used to calculate enthalpy and mass flow rate using thermodynamic relationships and two-phase flow correlations. This paper presents the field validation of that model through a new series of experiments conducted in 2025 at Theistareykir and Krafla geothermal fields in Iceland. The model was applied in real-time to differential pressure data collected across different operating conditions. At the same time, measurements using the water tracer dilution method were taken from the same wells to provide reference values to validate the models' predictions. The validation focuses on model performance across different well types and flow regimes. Evaluation metrics such as RMSE, and MAPE are used to evaluate the accuracy of model predictions against reference measurements. Preliminary results show the model achieves high accuracy in steam quality prediction (RMSE ≈ 0.011), supporting reliable estimation of enthalpy and flow rate in real time. This work aims to bridge the gap between academic development of machine learning models and their practical implementation. By validating the system under real field conditions, we demonstrate the potential of machine learning to improve reservoir monitoring, reduce operational costs, and enable continuous well performance tracking without interrupting production. |