Soil nitrogen (N) loss is a part of N biogeochemical processes, which plays an important role in the agricultural, ecological and environmental management. Because it is difficult to assess the temporal and spatial changes of different N forms in leachates by field measurement methods, conceptual and physical models are usually used to describe soil N loss. However, soil N models are often associated with multiple sources of uncertainty (e.g., model parameter and structure), which may largely influence the reliability and accuracy of the models. The multimodel ensemble prediction (MEP) is specifically designed to reduce the parameter and structural uncertainty in N biogeochemical modelling by representing a set of candidate models. However, the existing MEP methods still need to be improved by integrating various kinds of prior knowledge and quantifying each part of predictive uncertainty. In addition, published studies mainly focused on the regional scale MEP of the land carbon balance. However, the regional scale MEP of soil N losses is lacking. This paper firstly proposed the MEP methods of soil N losses at different spatial scales: 1) using the Monte-Carlo sampling to randomly alter the soil and crop parameters governing the N cycle and driving multiple soil N models at plot scale; and 2) generating an ensemble of TIGGE (THORPEX Interactive Grand Global Ensemble) weather forecasts and an ensemble of random soil and crop parameters and driving multiple soil N models at regional scale. This study also discussed different methods used for realizing MEP. It is found that the ensemble mean produced a large bias when simulating soil N losses. By using the bias correction technique, the RMSEs of the ensemble mean decreased by 57.5%~86.0%. Overall, the MEP can enhance our understanding of soil N cycle. In addition, this study is also helpful to accurately estimate the response of soil N loss to global change and provide support for agricultural production and environmental protection.

土壤氮（N）损失是氮生物地球化学过程的一部分，在农业、生态和环境管理中发挥着重要作用。由于很难通过现场测量方法评估渗滤液中不同形态氮的时空变化，因此通常使用概念模型和物理模型来描述土壤氮流失。然而，土壤氮模型通常与多种不确定性来源（例如模型参数和结构）相关联，这可能在很大程度上影响模型的可靠性和准确性。多模型集合预测 (MEP) 专门用于通过表示一组候选模型来减少 N 生物地球化学建模中的参数和结构不确定性。然而，现有的 MEP 方法仍然需要通过整合各种先验知识和量化每个部分的预测不确定性来改进。此外，已发表的研究主要集中在土地碳平衡的区域尺度 MEP。然而，缺乏土壤氮流失的区域尺度 MEP。本文首次提出了不同空间尺度土壤氮素损失的MEP方法：1）利用蒙特卡罗采样随机改变控制氮循环的土壤和作物参数，并在小区尺度上驱动多个土壤氮模型；2) 生成 TIGGE（THORPEX Interactive Grand Global Ensemble）天气预报和随机土壤和作物参数的集合，并在区域尺度上驱动多个土壤 N 模型。本研究还讨论了用于实现 MEP 的不同方法。发现在模拟土壤 N 损失时，整体均值产生了很大的偏差。通过使用偏差校正技术，集合均值的RMSE降低了57.5%~86.0%。总体而言，MEP 可以增强我们对土壤氮循环的理解。此外，本研究也有助于准确估计土壤氮流失对全球变化的响应，为农业生产和环境保护提供支持。