Existing thunderstorm identification methods typically generate two-dimensional composite reflectivity data based on raw radar data, neglecting the three-dimensional hierarchical structure information inherent in the original radar data. To address this issue, a method for identifying thunderstorm areas using three-dimensional volume radar echo data is proposed. This method processes the three-dimensional volume radar data derived from the original radar data through a three-dimensional convolutional neural network, retaining and utilizing the inter-layer structural information of the original data to enhance the accuracy of thunderstorm identification. An independently designed global feature extraction module is introduced into the three-dimensional UX-Net architecture, and group normalization is used instead of the traditional normalization method, enhancing the model's ability to extract global features and improving the stability of training with small batch sizes. The model was tested using SWAN radar echo data and ground flash location data provided by the Hunan Provincial Meteorological Bureau, achieving optimal performance in the Dice coefficient comparison. Experimental results show that the model demonstrates better effectiveness and accuracy in thunderstorm area identification.