To address the strong non-stationarity, pronounced abruptness, and limited uncertainty quantification of traditional point prediction methods for microseismic energy series in deep mines, a probabilistic interval prediction method for microseismic energy was proposed based on Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) and a hybrid TCN-BiLSTM-Attention network. Microseismic monitoring data collected from a deep metal mine were used for model validation. Microseismic energy, seismic moment, apparent stress, and inter-event time interval were selected as the original input features. The microseismic energy series was decomposed into multi-scale components by ICEEMDAN, and the obtained intrinsic mode function components were fused with the original source parameters to construct the input feature matrix. Local temporal features were extracted using a Temporal Convolutional Network (TCN), bidirectional temporal dependencies were captured using a Bidirectional Long Short-Term Memory network (BiLSTM), and the adaptive representation of critical historical time steps and important features was enhanced by a multi-head self-attention mechanism. Meanwhile, the predictive output was characterized by a Gaussian distribution, and the model was trained using the Continuous Ranked Probability Score (CRPS), thereby enabling both point prediction and probabilistic interval prediction of microseismic energy. The results show that the proposed model achieved high prediction accuracy on the test set, with an R2 of 0.926, a 95% prediction interval coverage probability of 0.863, and a mean prediction interval width of 2.195, indicating favorable interval prediction performance and uncertainty characterization capability. Further analysis showed that the predicted standard deviation output by the model was significantly associated with high-energy extreme events, which could provide probabilistic support for the early warning of dynamic disasters. The interpretability analysis indicated that the high-frequency components obtained by ICEEMDAN made the greatest contribution to model prediction, and that the multi-head self-attention mechanism could effectively focus on critical temporal information. The results provide a probabilistic decision-making reference for microseismic monitoring, early warning, and intelligent prevention and control of dynamic disasters in deep mines.