Mechanism-informed data-driven forecasting of roof cavities in longwall mining

Roof cavity is a typical form of longwall face instability arising from complex geological, geotechnical, and roof-shield interactions, posing risks to operational safety and productivity. This study develops a mechanism-informed data-driven approach for cavity forecasting by integrating numerical modelling with machine learning (ML). Discontinuum modelling combined with a Taguchi L9 sensitivity analysis identifies cover depth as the most influential factor, followed by mining height and roof strata competency. The modelling results also indicate an initial increase in shield leg pressure followed by pronounced unloading as the cavity develops, supporting the use of shield pressure drop as an indicator of cavity occurrence. Based on this mechanistic interpretation, an AutoEncoder-Long Short-Term Memory (AE-LSTM) architecture is developed to capture spatiotemporal patterns in shield monitoring data. The ML-based model exhibits temporal prediction accuracies of 89.3% to 94.7% and three-zone spatial localisation accuracies of 81.7% to 88.9% under within-panel validation, enabling one-cut-ahead forecasting and regional identification of cavity-prone areas along the face. Cross-site validation using independent datasets from another mine yields accuracies of 83.9% to 89.0% and F 1 -scores of 79.0% to 86.6%, demonstrating the transferability of the forecasting approach. The probability outputs are further interpreted as low, moderate, and high cavity risk levels to offer graded risk information beyond binary cavity or non-cavity classification. An online tool is developed to integrate data input, model execution, forecast visualisation, and risk interpretation, providing a practical basis for early warning and targeted roof management in longwall mining.