Abstract: To address the issues of low prediction accuracy, insufficient capability to capture mutation characteristics, and poor generalization under complex working conditions caused by the strong volatility, high nonlinearity, and significant mutability of mine power loads, a high-precision and robust deep learning-based prediction model for mine power loads is developed in this study. Taking the convolutional neural network and bidirectional long short-term memory network (CNN-BiLSTM) as the basic framework, the model first introduces the maximal information coefficient for adaptive screening of input features, and enhances the model’s perception of key features by integrating the squeeze-and-excitation mechanism. Subsequently, part of the LSTM units in BiLSTM are replaced with gated recurrent units, and residual connections are incorporated to optimize the gating structure while alleviating the gradient vanishing problem in long time-series modeling. Finally, Bayesian optimization is employed to achieve global optimization of model hyperparameters, avoiding the subjective dependence of manual parameter tuning. Experimental results demonstrate that the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) of the proposed model for mine power load prediction are 19.42 kW, 26.26 kW, and 2.23%, respectively, with an R2 value as high as 0.983. The relative prediction error at mutation nodes is 3.57%, the single-round training time is 20.8 s, and the single prediction time is 118.6 ms, all of which are superior to those of existing models. This study provides a novel method for the accurate prediction of mine power loads and exhibits important engineering application value for promoting the safe, efficient, and low-carbon development of intelligent mines.