Temporal Fusion Transformer for Fracture Evolution Prediction in Hot Dry Rock Hydraulic Fracturing

Hot dry rock (HDR) represents a highly promising sustainable clean energy resource for the future, with the fracture network induced by hydraulic fracturing serving as the cornerstone for thermal energy extraction efficiency. However, conventional numerical simulation tools are computationally expensive and fail to meet the requirements for real-time, dynamic on-site predictions. While machine learning models, such as traditional artificial neural network (ANN) and Long Short-Term Memory (LSTM), suffer from severe “black-box” limitations, they also lack the capability to capture the dynamic evolutionary characteristics of complex time series. To resolve these limitations, this study proposes a deep learning framework based on the Temporal Fusion Transformer (TFT) to dynamically predict the temporal evolution of fracture morphology. The research dataset was automatically generated in batches using GOHFER software (version 9.5.6), and feature selection was subsequently completed through parametric sensitivity analysis. A comparative analysis between the TFT model and traditional baseline models (ANN and LSTM) demonstrates that the TFT model achieves superior accuracy in capturing non-linear features that evolve dynamically over time, effectively overcoming the severe underfitting issues exhibited by conventional models under complex temporal constraints. Furthermore, leveraging the inherent attention mechanism of the TFT, this study elucidates the influence weights of injection rate, proppant concentration, and carrier fluid volume on fracture propagation across different fracturing stages, thereby enhancing the interpretability of the deep learning model. This research provides a novel decision-making tool that balances high efficiency with interpretability for the efficient and sustainable design of HDR.