Large-scale photovoltaic (PV) deployment in arid deserts may alter land–atmosphere interactions and influence groundwater systems, yet such impacts remain poorly quantified due to limited high-resolution observations. To overcome the coarse spatial resolution of GRACE data, this study develops a CNN-LSTM-Attention deep learning framework to downscale terrestrial water storage anomalies (TWSA) from 0.25° × 0.25° to 0.1° × 0.1° over the Huadian PV base in the Tengger Desert, China, during 2004–2024. Groundwater storage anomalies (GWSA) were derived using a water-balance approach, and piecewise linear regression was applied to detect trend shifts associated with PV development. Results show a persistent decline in TWSA and GWSA before 2022, followed by short-term recovery signals afterward. Groundwater responses exhibit greater magnitude and delayed behavior relative to soil moisture. Spatial analysis reveals stronger variability and more frequent deficits in the western subregion, indicating intra-base heterogeneity. A seasonal phase analysis identifies an approximately six-month lag between soil moisture and groundwater, highlighting constraints from deep vadose-zone processes. The findings suggest that groundwater dynamics reflect the combined effects of climate variability, infiltration lag, and PV-related land surface modification rather than a single driver. This study demonstrates the potential of deep-learning-based GRACE downscaling for groundwater monitoring in human-modified arid regions and provides insights for sustainable water management under renewable energy development.
Chen et al. (Thu,) studied this question.