Advancing fractured geothermal system modeling with artificial neural network and bidirectional gated recurrent unit

Geothermal energy plays a pivotal role in the global energy transition towards carbon-neutrality, providing a sustainable, renewable, and abundant source of clean energy in the fight against climate change. Despite advancements, the optimal engineering of geothermal systems and energy extraction remains challenging, particularly in accurately predicting production temperatures. Here, we present an innovative numerical approach using a hybrid neural network that merges Artificial Neural Network (ANN) and Bidirectional Gated Recurrent Unit (BiGRU). With this hybrid network, we comprehensively assess 22 influential factors, including construction parameters, physical parameters, and well layout, which influence thermal breakthrough time and production temperature across varying fracture density. While the ANN captures the nonlinear interplay between static constraints and thermal breakthrough time, the BiGRU adeptly handles the temporal intricacies of production temperature. We examine the impact of ANN parameters on model performance, in comparison with conventional temporal models like Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gate Recurrent Unit (GRU), and BiGRU. Our findings reveal that augmenting hidden layers and neurons in ANN enhances its capacity to model intricate nonlinear processes, albeit with a risk of overfitting. Notably, the relu activation function emerges as optimal for managing nonlinear processes, while BiGRU excels over RNN, GRU, and LSTM models in forecasting production temperature of fractured geothermal systems, owing to its ability to extract implicit information from time series across historical and future trajectories. Crucially, the prediction uncertainty, measured by Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), remains within 0.15, underscoring the precision and efficacy of our hybrid approach in forecasting geothermal energy extraction. This study presents a significant stride towards a high-precision and efficient predictive framework crucial for advancing geothermal energy extraction in the broader context of renewable energy transition endeavors.