The Gravity Recovery and Climate Experiment (GRACE) mission ended its operation in October 2017, and the GRACE Follow‐On mission was launched only in May 2018, leading to approximately 1 year of data gap. Given that GRACE‐type observations are exclusively providing direct estimates of total water storage change (TWSC), it would be very important to bridge the gap between these two missions. Furthermore, for many climate‐related applications, it is also desirable to reconstruct TWSC prior to the GRACE period. In this study, we aim at comparing different data‐driven methods and identifying the more robust alternatives for predicting GRACE‐like gridded TWSC during the gap and reconstructing them to 1992 using climate inputs. To this end, we first develop a methodological framework to compare different methods such as the multiple linear regression (MLR), artificial neural network (ANN), and autoregressive exogenous (ARX) approaches. Second, metrics are developed to measure the robustness of the predictions. Finally, gridded TWSC within 26 regions are predicted and reconstructed using the identified methods. Test computations suggest that the correlation of predicted TWSC maps with observed ones is more than 0.3 higher than TWSC simulated by hydrological models, at the grid scale of 1° resolution. Furthermore, the reconstructed TWSC correctly reproduce the El Nino‐Southern Oscillation (ENSO) signals. In general, while MLR does not perform best in the training process, it is more robust and could thus be a viable approach both for filling the GRACE gap and for reconstructing long‐period TWSC fields globally when combined with statistical decomposition techniques.

中文翻译：

---

数据驱动技术的比较重建（1992-2002）和预测（2017-2018）使用气候输入的类似GRACE的网格总蓄水量变化

重力恢复和气候实验 (GRACE) 任务于 2017 年 10 月结束运行，GRACE 后续任务仅在 2018 年 5 月发射，导致大约一年的数据缺口。鉴于 GRACE 型观测仅提供对总储水量变化 (TWSC) 的直接估计，弥合这两个任务之间的差距将非常重要。此外，对于许多与气候相关的应用，在 GRACE 周期之前重建 TWSC 也是可取的。在这项研究中，我们的目标是比较不同的数据驱动方法并确定更可靠的替代方案，用于在间隙期间预测类似 GRACE 的网格 TWSC，并使用气候输入将它们重建到 1992 年。为此，我们首先开发了一个方法框架来比较不同的方法，例如多元线性回归 (MLR)、人工神经网络 (ANN) 和自回归外生 (ARX) 方法。其次，开发指标来衡量预测的稳健性。最后，使用识别的方法预测和重建 26 个区域内的网格 TWSC。测试计算表明，在1°分辨率的网格尺度下，预测的TWSC图与观测图的相关性比水文模型模拟的TWSC高0.3以上。此外，重建的 TWSC 正确地再现了厄尔尼诺-南方涛动 (ENSO) 信号。一般来说，虽然MLR在训练过程中表现不佳，