Abstract Multi-modelling aims to make use of the strengths of single hydrologic models to improve the accuracy of simulating the watershed system behavior. Considering hydrological signatures such as the flow duration curve segmentation in the calibration of each hydrologic model leads to a better parameter identifiability. In this study, a novel weighted average model-wrapper based on flow duration curve segmentation is introduced to aggregate the calibrated models into a multi-model. The proposed framework is applied to develop a model-wrapper of the Upper Assiniboine River Basin for flood forecasting upstream of the Shellmouth reservoir in the Prairie region of Canada. The HEC-HMS, HBV-EC, HSPF, and WATFLOOD hydrologic models that are being used at the Hydrologic Forecast Centre of Manitoba Infrastructure for operational inflow forecasting are calibrated using signature-based multi-objective optimization. These models have significantly different structural complexities. The calibration of each of these models is set up as three simulation-optimization problems with different objective functions to balance the model capability in simulating multiple important hydrological signatures. Results show that the model-wrapper outperforms each of the single calibrated models that are of operational use at Manitoba Infrastructure, e.g. NSE improved from 0.44 for the best individual model to 0.76 for the model-wrapper in the calibration period. Moreover, the weights associated with each hydrologic model component indicate the contribution rate of the individual models to the model-wrapper in high-flow, mid-flow, and low-flow portions of streamflow time series. Quantifying the contribution of each model component provides a deeper insight into model selection strategy, especially when a component has minimal or no contribution, e.g. HEC-HMS and HBV-EC in this paper, to the model-wrapper performance in all ranges of streamflow simulation compared to other model components.

中文翻译：

---

基于特征的水文模型多建模和多目标校准：在加拿大草原洪水预报中的应用

摘要 多重建模旨在利用单一水文模型的优势，提高流域系统行为模拟的准确性。在每个水文模型的校准中考虑水文特征，例如流量持续时间曲线分段，导致更好的参数可识别性。在本研究中，引入了一种基于流量持续时间曲线分割的新型加权平均模型包装器，将校准后的模型聚合为多模型。所提议的框架用于开发上阿西尼博因河流域的模型包装器，用于加拿大草原地区壳茅斯水库上游的洪水预报。HEC-HMS、HBV-EC、HSPF、和 WATFLOOD 水文模型正在曼尼托巴基础设施水文预测中心用于业务流入预测，使用基于特征的多目标优化进行校准。这些模型具有显着不同的结构复杂性。每个模型的校准被设置为三个具有不同目标函数的模拟优化问题，以平衡模型模拟多个重要水文特征的能力。结果表明，模型包装器的性能优于曼尼托巴基础设施运营中使用的每个单一校准模型，例如，在校准期间，NSE 从最佳个体模型的 0.44 提高到模型包装器的 0.76。而且，与每个水文模型组件相关联的权重表示各个模型在水流时间序列的高流量、中流量和低流量部分对模型包装器的贡献率。量化每个模型组件的贡献可以更深入地了解模型选择策略，尤其是当组件对所有流模拟范围内的模型包装器性能贡献最小或没有贡献时，例如本文中的 HEC-HMS 和 HBV-EC与其他模型组件相比。