In this study, we developed and evaluated a hybrid framework for reservoir inflow forecast. This framework is unprecedented, which integrates new quasi-globally available observation-, satellite-, or model-based datasets using machine learing models to forecast inflow at the local scale. Under this framework, we compared random forests, gradient boosting machine, extreme learning machine, M5-cubist, elastic net, as well as their multi-model ensemble using Bayesian model averaging (BMA), and tested contributions from different input datasets, including retrospective forecast (reforecast) from florecast oriented low ocean resolution (FLOR) dynamic climate model, satellite-based hydrologic products, and climate phenomenon information. The performance was evaluated using Kling–Gupta efficiency (KGE) and correlation coefficient (R) in two headwater reservoirs, Harris reservoir in the humid Alabama–Coosa–Tallapoosa river basin and the Navajo reservoir in the arid Upper Colorado River Basin (UCRB). The results showed that for the Harris reservoir, the BMA combining five models with antecedent inflow and satellite-based hydrologic information as model inputs provided the best performance (KGE = 0.66, R = 0.76). For the Navajo reservoir, the gradient boosting machine model with all variables combined as input showed the best performance (KGE = 0.76, R = 0.83). Satellite-based soil moisture and evaporation consistently showed significant contributions to the inflow forecast. Benefits from climate indices and FLOR reforecast varied by locations, with more benefits coming from climate indices than FLOR potential evaporation reforecast at the Navajo reservoir in UCRB. Given the global coverage of the model inputs, our approach can be potentially applicable to improve reservoir inflow forecasts in different regions of the world.

在这项研究中，我们开发并评估了用于储层流入预测的混合框架。这个框架是空前的，它使用机器学习模型集成了新的准全球可用的基于观测，卫星或模型的数据集，以预测本地规模的流入量。在此框架下，我们使用贝叶斯模型平均（BMA）比较了随机森林，梯度提升机，极限学习机，M5立方体专家，弹性网及其多模型集合，并测试了包括回顾性在内的不同输入数据集的贡献根据面向花播的低海洋分辨率（FLOR）动态气候模型，基于卫星的水文产品和气候现象信息进行预测（重新预测）。使用Kling–Gupta效率（KGE）和相关系数（R）对两个水源水库（湿润的阿拉巴马州–库萨–塔拉波萨河流域的哈里斯水库和干旱的上科罗拉多河流域（UCRB）的纳瓦霍水库）进行了性能评估。结果表明，对于哈里斯水库，BMA将五个模型与先入流和基于卫星的水文信息相结合作为模型输入，可提供最佳性能（KGE = 0.66，R = 0.76）。对于纳瓦霍水库，将所有变量组合为输入的梯度提升机模型显示出最佳性能（KGE = 0.76，R = 0.83）。卫星为基础的土壤水分和蒸发量一直对入流预报显示出重大贡献。气候指数和FLOR的预测收益因位置而异，与气候变化指标相比，UCRB纳瓦霍水库重新预测的FLOR潜在蒸发量带来的收益更多。考虑到模型输入的全球覆盖范围，我们的方法可能适用于改善世界不同地区的储层流入预测。