Research on Real-Time Hydraulic Fracturing Prediction and Evaluation Model Based on Neural Networks

ABSTRACT: This study proposes a machine learning-based surrogate model for real-time hydraulic fracturing monitoring, overcoming traditional models' limitations in timeliness and small-sample applications. A numerical simulation framework integrating GOHFER software with automated Python scripts generated 12,000 fracture geometry datasets through parametric modeling. Parametric sensitivity analysis identified critical inputs including reservoir properties and fracturing fluid parameters. The preprocessing pipeline employed Isolation Forest for outlier removal and min-max normalization for feature scaling. An optimized neural network architecture with dual hidden layers was trained using NAdam optimizer and MAPE loss function, achieving efficient convergence. For temporal feature extraction from pumping data, a hybrid ANN-LSTM architecture was developed to enable real-time fracture propagation analysis and Stimulated Reservoir Volume (SRV) estimation. Experimental results demonstrate the model's sub-10-second response time with enhanced prediction accuracy, representing 85% efficiency improvement over conventional methods. The integrated methodology establishes a novel workflow combining automated numerical simulation with deep learning, providing an effective solution for dynamic fracture diagnostics and operational optimization in unconventional reservoirs. This data-driven approach significantly enhances fracturing treatment evaluation while reducing computational costs associated with physics-based simulations.