Risk identification and early warning method of old goaf floor Ordovician limestone water inrush based on attention-LSTM and isolation forest

Abstract

Mine water inrush from the Ordovician limestone aquifer is a major coal-mine safety threat, especially beneath old goaf floors where precursor signals are weak and concealed. This study develops an intelligent early-warning method integrating an Attention-LSTM network and an Isolation Forest model. Five-minute borehole water-level data were collected at Xin'an Mine in Hebei, China, from October 2020 to December 2021. Data from four boreholes (XO5, KM, O2–4, and O2–6) were used. For each borehole, the previous 60 water-level observations were input to predict the next-hour water level. The Attention-LSTM model was trained on stable-period data to establish a dynamic baseline of normal water-level variation. Prediction residual features, including absolute error (AE) and its first-order difference (ΔAE), were then extracted as inputs to the Isolation Forest model for unsupervised anomaly detection. The model achieved MAE values of 0.170, 0.169, 0.028, and 0.011 m and NSE values of 0.971, 0.970, 0.973, and 0.994 for XO5, KM, O2–4, and O2–6, respectively. The AE–ΔAE feature combination produced the best anomaly-detection performance, and the multi-borehole warning strategy successfully identified a high-risk state approximately 3.5 h before the actual field water-inrush event.