Evaluation of Global Arps and Machine Learning Models for Predicting Geothermal Water Production of New Bakken Wells

ABSTRACT: This study evaluates the principle of analogy for predicting warm produced water rates of wells drilled within the Bakken Formation. 40 horizontal wells completed in the Bakken pool, located across different fields and with production records spanning 2009–2025, were analyzed. 35 wells were used as training wells to develop two global predictive models: (1) an analytical Arps decline model optimized using least squares regression, and (2) a global Long Short-Term Memory model trained using the Adam optimizer. The trained global models were then applied to predict water production rates for 5 unseen wells, treated as newly drilled wells. The global Arps decline model was able to qualitatively capture the decline trend of four out of the five test wells, yielding R2 values between 0.582 and 0.870 and Mean Absolute Percentage Error ranging from 26% to 48.7%. The relatively high MAPE values are attributed to abrupt production fluctuations present in the field data. The global LSTM model, implemented with a six-month sequence length, demonstrated improved performance in capturing water production trends by achieving R2 values between 0.752 and 0.958 across the five unseen wells, and MAPE between 12.9% – 14.3%. The results indicate that while a global analytical decline model provides reasonable first-order predictions under the principle of analogy, sequence-based machine learning models offer stronger capability in modeling nonlinear and time-dependent behavior in produced water rates.