Abstract: Rockburst is particularly hazardous, rendering the accurate prediction of rockburst risks one of the urgent challenges in this field of research. This study proposes and establishes a long short-term memory (LSTM) rockburst risk prediction model (CAL model), incorporating the attention mechanism with convolutional neural networks (CNNs). Specifically, we compared and tested 4 feature selection methods using rockburst sample data collected from both domestic and international sources; constructed the CAL model by combining CNNs, LSTM and the attention mechanism, followed by training and prediction; performed a ten-fold cross-validation, comparing the results with K-nearest neighbors (KNN), support vector machine (SVM) and deep neural network (DNN) models, alongside ablation experiments involving the CNN, LSTM, and CNN-LSTM models. The experimental results show that the CAL model exhibited a high prediction accuracy of 98.6%, precision of 98.8%, recall of 98.6%, and F1-score of 0.987, with all 4 evaluation metrics outperforming existing mainstream models. Finally, an on-site practical engineering inspection was conducted, proving that the CAL model can effectively predict rockburst risk.