The Huang-Huai-Hai River basin of China suffers from severe water scarcity during recent decades under dual impacts from climate change and human activities. Quantifying the change of terrestrial water resources in the region as well as its driven factors is of significant to understand hydrology process and develop sustainable water policy. This study proposes a machine- learning model for reconstructing past changes in terrestrial water storage (TWS) in the Huang-Huai-Hai River basin, China. The model is trained with the observations of changes in TWS from the Gravity Recovery and Climate Experiment mission (GRACE) satellites and climate and human water use datasets at monthly scale. The trained model is used to reconstructed historic TWS. The changes of the reconstructed TWS are compared with results of a set of global land surface and hydrology models as well as in-situ soil moisture measurements. Moreover, the contribution of climatic and human factors to the machine learning model are quantified. Results show that the presented approach generally outperforms the global land surface models and hydrological model examined in this study. It well reproduces the spatial patterns of GRACE observations and reconstructs past changes in TWS that consistent with GRACE observations. Climate variables dominate the importance rankings in the machine learning model, and prior one-month precipitation has the highest importance value. The model including human interventions performs better than that without human interventions, and the irrigation, industry, and domestic water uses contribute equally to the model. This study provides a flexible and easily implementable model to bridge the gap between GRACE observations and the past change in TWS. The model is applicable in the area with intense human activities and the results have potential to enrich and be assimilated into hydrological models.