Precisely quantifying carrying capacity of water resource (CCWR) can address the imbalance between water supply and water demand thus, and promote socio-economic and environmentally sustainable development of water-scarce regions. Through coupling input-output analysis with deep learning methods, an input-output CNN-LSTM-Attention (abbreviated as IOCLA) model is developed. IOCLA has advantages in identifying the key drivers of virtual water, predicting future water supply and water demand, and shrinking water shortage under socioeconomic and climate scenarios (SSPs-RCPs). Methodologically, IOCLA also excels at tracing economy-wide virtual-water transfers along supply chains and capturing nonlinear spatiotemporal dynamics. IOCLA is then applied to the Inner-Shaan-Ning region (a typical arid region located in the upper reaches of the Yellow River Basin), which economic development is constrained by the shortage of water resources. The main findings are summarized as follows: (i) in the Inner-Shaan-Ning region, the main driving factors of virtual water are water intensity, investment, exports, and imports; (ii) water shortages are lowest under SSP126 and highest under SSP585 (e.g., by 2050, water shortages rank as SSP126 < SSP245 < SSP370 < SSP585); (iii) compared with SSP245, SSP370, and SSP585, the CCWR under SSP126 increases by 6.7%, 23.1%, and 45.4%, respectively, indicating that SSP126 scenario is more suitable for this region. The findings are conductive to adjusting industrial structure, reducing the imbalance between water supply and demand, and facilitating socio-economic and environmentally sustainable development.