Groundwater in arid regions faces critical quantity-quality crises exacerbated by agricultural expansion, urbanization, and climate change, threatening water security and ecological stability, particularly in endorheic basins like the Minqin Basin. This study employs numerical modeling, data-driven, and risk evaluation approaches to analyze the spatiotemporal distribution characteristics of groundwater and evaluate overexploitation risks. The results show that the groundwater system exhibited a negative water budget during 2008-2014 but transitioned to a positive water budget from 2015 to 2020. Comprehensive evaluation via goodness-of-fit metrics and model mechanism analysis demonstrated that the convolutional neural networks and long short-term memory (CNN-LSTM) model, confirming its superior simulation performance in the study area. Groundwater overexploitation exhibits a three-dimensional evolution characteristic of "area expansion, rate attenuation, and weakening of abnormal zones". The CNN-LSTM model provides a reliable tool for regional water resource early warning. This study provides an integrated monitoring-simulation-warning-assessment management framework for global arid regions with analogous conditions, offering direct practical reference for irrigated agricultural areas addressing water sustainability challenges. Findings will advance understanding of groundwater cycling in arid regions and provide a theoretical foundation for sustainable groundwater resource management in the Minqin Basin.