Abstract: In-depth research on the rock burst early-warning under hard roof geological conditions was conducted by integrating deep learning techniques. A time-series rock burst early-warning method was proposed based on the fusion of explicit physical warning indicators and data-driven implicit temporal warning indicators, and the evolution patterns of multiple physical warning indicators prior to high-energy microseismic events were investigated. The Genetic Algorithm-Optimized Backpropagation Neural Network (GA-BP) was used to screen four physical warning indicators suitable for the hard roof working face: microseismic activity quiescence β_n, dynamic parameter of small earthquakes P(b) microseismic activity index A(b), and empirical constant in the G-R relation b. A Convolutional Neural Network combined with t-Distributed Stochastic Neighbor Embedding (CNN-T-SNE) was employed to extract temporal warning indicators from microseismic and support resistance data. Deep learning algorithms, including the Nutcracker Optimization Algorithm (NOA) and Long Short-Term Memory (LSTM) networks, were applied to establish a time-series early-warning model for hard roof-type rock bursts, and the model's predictive performance was validated and analyzed through case studies. The results show that: GA-BP-based feedback-iterative optimization screened β_n, P(b), A(b), and b as physical warning indicators with prediction F₁-scores of 0.88 and RMSE of 0.038, exhibiting strong sensitivity for impact warning. Before high-energy events, b, P(b), and β_n values were at low levels or declined rapidly, while only A(b) remained at or near its maximum. The CNN-T-SNE algorithm effectively extracted temporal warning indicators, with T-SNE providing clear visualization and separation of different risk levels, improving prediction accuracy. The temporal feature indicators achieved an F₁-score of 0.91 and RMSE of 0.021, while microseismic and support resistance indicators achieved F₁-scores of 0.80 and RMSE of 0.085, representing a 12% improvement in warning accuracy. The NOA algorithm optimized LSTM hyperparameters, yielding optimal values of learning rate 0.033, batch size 8, and Dropout rate 0.2. The LSTM-NOA network, trained on both physical and temporal warning indicators with a one-day prediction horizon, achieved an F₁-score of 0.96, representing a 0.04 improvement in F₁-score and a 4% increase in warning accuracy compared to traditional LSTM methods, with enhanced sensitivity to high-energy events.