Abstract: In contemporary coal mining, multi-level mining has become a prevalent mode of operation. However, the challenges posed by complex geological structures and interactions among multiple aquifers have rendered accurate prediction of mine water inflow a significant industrial challenge. The variability in mine water inflow directly impacts the safety of underground workers, the stability of equipment operation, and mining efficiency. Substantial deviations in prediction outcomes may lead to severe safety incidents, such as water inrush events. Consequently, the exploration of efficient and precise methods for predicting mine water inflow is critical for ensuring safe production and achieving sustainable resource development. To address these issues, this paper employs deep learning algorithms to analyze collected water inflow data, uncovering intricate relationships and patterns within the dataset. Leveraging the robust time-series data processing capabilities of Long Short-Term Memory Network （LSTM）, the model effectively captures long-term dependencies in water inflow variations over time. Additionally, the feature extraction advantages of Convolutional Neural Network （CNN） are utilized to mine local features from the water inflow data. Furthermore, the Attention Mechanism （Attention） is incorporated to focus the model on key factors influencing water inflow changes, leading to the construction of an LSTM-CNN-Attention model. This paper uses a specific coal mine in North China as an engineering case, systematically collects real-time water inflow data from the 7-12 levels of the mine from January 2016 to June 2023. The dataset encompasses multi-dimensional information, including different mining stages and seasonal variations. The data are scientifically partitioned into training, validation, and test sets in a ratio of 7∶2∶1. A gradient descent algorithm with excellent convergence properties is employed to optimize network parameters and regularization terms, ensuring the model’s stability and generalization capability. To validate the model’s superiority, it is compared and analyzed against traditional BP neural network, Autoregressive Integrated Moving Average （ARIMA）, and single Long Short-Term Memory Network （LSTM） models. The research results indicate that the LSTM-CNN-Attention model has a mean absolute error （MAE） of 0.121 0, a mean square error （MSE） of 0.023 0, a root mean square error （RMSE） of 0.151 0, and a coefficient of determination （R²） of 0.975 0. Compared to benchmark models, it demonstrates superior generalization ability and prediction accuracy. By innovatively integrating multimodal deep learning algorithms, the LSTM-CNN-Attention model provides an advanced and efficient approach for prediction of mine water inflow.