Background Mine resistivity prediction serves as a core technique for water hazards monitoring in coal mines. However, due to sparse monitoring points and insufficient spatial resolution, conventional prediction approaches fail to capture dynamic resistivity variations under complex geological conditions.

Objective and Methods  To address this challenge, this study developed a deep learning-based time series forecasting model integrating spatial features for inter-borehole resistivity. Using a prediction framework constructed based on long short-term memory (LSTM) and critical spatial monitoring points selected through Pearson correlation analysis, this model (also referred to as the LSTM model) enabled high-precision prediction of resistivity in unknown areas. Specifically, the electrical connectivity between monitoring points was quantified using spatial correlation heatmaps. The optimal number of neighboring points was determined at five to reduce data redundancy and enhance the generalization capability of the LSTM model. The model architecture captured temporal dependencies in resistivity data using the gating mechanisms of LSTM and incorporated spatial information from multiple monitoring points, reducing the sensitivity of conventional models to vanishing gradients associated with long sequences. Against the engineering background of mining face 61304 in the Tangjiahui coal mine, Ordos City, this study compared the prediction performance of the LSTM model and the recurrent neural network (RNN) model based on the monitoring data from directional boreholes in the coal seam floor.

Results and Conclusions The results indicate that the LSTM model outperformed the RNN model in terms of mean absolute error (E_MA = 0.0582), root mean square error (E_MS = 0.005 2), and coefficient of determination (R² = 0.937 7). Under 10% noise interference, the R² value of the LSTM model decreased by merely 0.02, demonstrating strong robustness. The LSTM model was applied to the dynamic monitoring during the early, middle, and late grouting stages. A high-density spatiotemporal resistivity dataset was obtained by supplementing sparse measured data with predicted data. The inversion imaging results based on this dataset confirm that the LSTM model can restore the continuity of geological structures that is ignored in low-density monitoring while successfully identifying weak anomaly areas that cannot be reflected in the inversion results derived using low-density data. The LSTM model can effectively overcome the limitations of low spatial resolutions in traditional monitoring, offering a robust technical approach for the dynamic monitoring and accurate early warning of water hazards under complex geological conditions.