Hybrid deep learning Approach for roof water inflow Prediction: Integration of microseismic energy density analysis with CNN-BiLSTM

Roof water inflow under thick aquifers is governed by strongly nonstationary coupling among mining-induced roof damage, water-conducting pathway development, and delayed aquifer response. Using 520 consecutive daily observations and 12,050 complete microseismic records from the 102 working face of the Jinjitan Coal Mine, this study developed a structure-informed convolutional neural network–bidirectional long short-term memory (CNN–BiLSTM) framework for next-day inflow prediction. A three-dimensional microseismic energy–density field was constructed on a 5 m voxel grid, projected into Y–Z slices, and used to derive a daily roof-damage index ( D d ). The results showed a clear four-stage evolution of high-energy microseismic activity, with Stage III representing the critical period of concentrated fracturing and rapid pathway development. During Stage III-2, D d increased persistently, whereas inflow fluctuations lagged by approximately 10–15 days. D d showed the strongest association with inflow (Mantel’s r ≈ 0.91, p ≤ 0.001). Under a strict chronological split, the proposed model achieved a mean absolute error (MAE) of 11.86 m 3 /h, a root mean square error (RMSE) of 14.77 m 3 /h, and a mean absolute percentage error (MAPE) of 3.24 %.