Hot dry rock (HDR) is a low-carbon geothermal resource and can be an alternative to fossil fuels. The prediction of sequential production is essential for operating a sustainable geothermal system, but its prediction using the numerical simulation method is time-consuming. Deep learning provides an alternative way for this task. However, previous studies mainly took the original production data as learning data, or more complexity, some constraint conditions were utilized in previous deep learning models to improve the prediction accuracy but enhance the process time. Besides, the research about comparisons of different deep learning algorithms lacks. In this work, three different deep learning algorithms, including the Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Transformer, are applied to forecast the productivities of a three-horizontal-well EGS. A difference value operation is worked on the data samples at adjacent time steps, followed by comparisons of model performance. The results show that a difference value operation on data samples promotes the predictive ability of a deep learning model compared to a model with the original data. According to the root mean squared errors (RMSE), mean absolute errors (MAE), and mean absolute percentage errors (MAPE), Transformer demonstrates the highest prediction accuracy with a 0.00005 MAPE value. Importantly, deep learning significantly save prediction time by 14 times compared to numerical simulation. This work proves that a deep learning model under a difference value operation can be an encouraging alternative to numerical simulation in predicting time series geothermal production. The proposed data operation method can provide a promising performance so that complex operations to improve the model performance like complex field constraints can be efficiently avoided. Furthermore, the proposed method can also be applied to predictions in other industries.

You can access this article if you purchase or spend a download.