Abstract: [Objective] Shear wave velocity is one of the key parameters characterizing the physical and mechanical properties of subsurface media and plays an important role in oil and gas reservoir evaluation. In carbonate fracture cave reservoirs, shear wave velocity is often difficult to measure directly due to drilling and logging limitations, and is therefore commonly predicted using rock physics models and empirical formulas. However, the complex structure and strong heterogeneity of carbonate fracture–cave reservoirs generally result in low prediction accuracy when using conventional methods. [Methods] This study presents a shear wave velocity prediction approach based on dimensionality reduction, reservoir type classification, and a Long Short-Term Memory (LSTM) neural network. First, undistorted logging curves are selected to correct distorted logging data to ensure input data quality. Eleven logging parameters, including acoustic time difference, density logging, and neutron logging, are reduced to five principal components using principal component analysis (PCA), effectively reducing data redundancy. Based on imaging logging responses and electrical characteristics, the reservoirs are classified into six types: dissolved pores, fractures, intact bedrock, unfilled caves, sandstone-filled caves, and gravel-filled caves. Support vector machine (SVM) is employed to perform reservoir type classification. [Results] On this basis, an LSTM neural network model is constructed to predict shear wave velocity for different reservoir types. Unlike traditional methods that rely on explicit modeling of fractures, pores, and caves, the proposed method directly uses logging curves and their principal components that are highly correlated with measured shear wave velocity, avoiding the need to construct complex rock physics models. Application of the proposed method to carbonate fracture cave reservoirs in the Tahe Oilfield shows that the predicted shear wave velocity has a high correlation with measured data, with correlation coefficients reaching up to 0.96 in fractured reservoirs. The overall prediction accuracy exceeds 91%, and the predicted shear wave velocity curves show good agreement with measured curves. [Conclusion] The results demonstrate that the proposed method provides an effective and efficient approach for predicting shear wave velocity in strongly heterogeneous carbonate fracture cave reservoirs and shows good application potential in ultra-deep carbonate reservoirs.