Rockburst　is　a　common　geological　disaster　in　underground　engineering,　posing　significant　challenges.　Data-driven　methods　provide　powerful　tools　for　the　assessment　and　early　warning　of　rockburst.　However,　existing　LSTM　and　its　derivative　architectures　face　a　trade-off　between　accuracy　and　stability.　We　propose　a　Hybrid　Convolutional　Long　Short-Term　Memory　Network　(HCLSTM)　to　address　this.　Unlike　previous　models,　HCLSTM　innovatively　integrates　onedimensional　and　two-dimensional　convolutions,　leveraging　their　strengths　in　short-term　feature　extraction　and　high-level　abstract　feature　extraction,　combined　with　the　long-term　memory　capability　of　LSTM.　This　architecture　achieves　both　high　accuracy　and　high　robustness.　To　fairly　evaluate　HCLSTM,　a　sliding　window　method　was　used　to　process　the　data　and　compare　it　with　single　LSTM　models　(Vanilla　LSTM,　Stacked　LSTM,　and　Bidirectional　LSTM)　and　convolutional　LSTM　hybrid　models　(CNN1d-LSTM　and　CNN2d-LSTM).　All　models　were　based　on　Optuna＇s　optimal　architecture.　The　results　showed　that　macro　metrics　(accuracy　and　Kappa　coefficient)　indicated　HCLSTM　as　the　best-performing　predictive　model,　while　micro　metrics　(Precision,　Recall,　and　F1)　demonstrated　their　superior　performance　in　predicting　minority　classes.　Additionally,　in　three　scenarios　exploring　the　impact　of　data　distribution,　data　partitioning　methods,　and　sequence　length　on　the　model,　HCLSTM　exhibited　the　highest　stability.　Overall,　experimental　results　indicated　that　HCLSTM　achieved　a　balance　between　accuracy　and　stability　when　handling　long　sequences　with　complex　spatiotemporal　characteristics,　outperforming　CNN1d-LSTM　and　CNN2d-LSTM.　This　provides　new　insights　for　the　extended　design　of　LSTM　models　and　offers　a　powerful　tool　for　the　analysis　and　evaluation　of　rockburst　in　engineering　applications.