To address the issue of inaccuracies in groundwater level predictions due to the insufficient consideration of groundwater-related factors, clustering methods for observation wells based on spatial distance, hydrogeological attributes, and a hybrid of distance and attributes were proposed. The significance of inter-well connectivity in groundwater level prediction was validated. Four models were designed, which were applied to simulate and predict groundwater levels in the karst water region of Jinan and compared with actual observations. The prediction results indicate that the combined model incorporating the connectivity characteristics of karst aquifers, known as convolution-long short-term memory (ConvLSTM), outperforms the traditional long short-term memory (LSTM) model. Among the models, the mix-multivariate-convolution-long short-term memory(M-MV-ConvLSTM) model, which accounts for wells of the same category based on the hybrid distance-attribute clustering results (characterized by strong connectivity), achieves the highest prediction accuracy and the smallest error. The average root mean square error is approximately 0.457, and the Nash-Sutcliffe efficiency is approximately 0.216, demonstrating a higher prediction accuracy than the traditional LSTM model. The research results is positioned to serve as a reference for real-time groundwater level prediction in karst regions.