Intelligent inversion method integrating broad learning and deep neural networks for in-situ stress prediction in strong compression tectonic zones

Abstract Accurate prediction of in-situ stress fields in intensely deformed geological settings represents a critical challenge for ultra-deep reservoir development. Conventional geomechanical approaches and existing deep learning methods face significant limitations in computational efficiency and prediction accuracy. This study introduces an intelligent inversion framework integrating broad learning system (BLS) with long short-term time-series network (LSTNet) to overcome these limitations. Our methodology introduces key innovations distinguishing it from conventional LSTM/BiLSTM approaches: a hybrid BLS-LSTNet architecture combining rapid feature expansion with sophisticated spatiotemporal modeling; streamlined workflow replacing computationally intensive pre-stack processing with efficient post-stack inversion enhanced by machine learning; and integration of physics-constrained seismic attributes. Validation in the Ketan thrust belt demonstrates 3-fold computational time reduction compared to conventional pre-stack methods, achieving 89% accuracy in blind-well tests for stress magnitude and orientation prediction. Comprehensive ablation studies confirm superior performance of our integrated architecture (R² = 0.85 for Vs prediction) over individual BLS (R² = 0.81), LSTNet (R² = 0.79), and baseline models including conventional LSTM (R² = 0.76). The methodology provides robust technical support for real-time drilling optimization in complex compressional environments, though performance decreases by 8%–12% in intervals with signal-to-noise ratio below 2.5. This work represents a significant advancement over existing approaches by effectively balancing computational efficiency with prediction accuracy for industrial-scale applications.