ABSTRACT With the increasing depth of coal mining, accurate prediction of periodic roof weighting is crucial for ground control and safety. Traditional methods based on data‐driven, which often treat sensor data as pure time‐series, are significantly affected by operational noise tied to the mining advance cycle and fail to account for spatial heterogeneity in pressure distribution. To address these limitations, this study proposes a novel region‐aware prediction model that integrates the region growing algorithm (RGA) with kernel density estimation (KDE)—the RGA‐KDE model. The model aligns analysis with the mining cycle by utilizing cycle‐end resistance, effectively filtering procedural noise. It dynamically partitions hydraulic supports into homogeneous regions based on spatial pressure trends using RGA and then applies KDE to model the probability distribution of stable intervals (advance distances between weighting events) within each region, enabling region‐specific probabilistic forecasting. Experiments demonstrate that the RGA‐KDE model significantly outperforms benchmark time‐series models (including ARIMA, LSTM, and their hybrids), achieving a substantially lower average prediction error and higher accuracy in rock pressure prediction. More importantly, the model successfully identifies regions with differing pressure behavior, with prediction accuracy particularly high in areas where the strata behavior exhibits more regular periodic intervals, highlighting its capability to adapt to spatial variations in ground pressure manifestation.
Weiqiang et al. (Sun,) studied this question.