当前位置: X-MOL 学术J. Hydrol. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
A Stacking Ensemble Learning Framework for Annual River Ice Breakup Dates
Journal of Hydrology ( IF 6.4 ) Pub Date : 2018-06-01 , DOI: 10.1016/j.jhydrol.2018.04.008
Wei Sun , Bernard Trevor

Abstract River ice breakup dates (BDs) are not merely a proxy indicator of climate variability and change, but a direct concern in the management of local ice-caused flooding. A framework of stacking ensemble learning for annual river ice BDs was developed, which included two-level components: member and combining models. The member models described the relations between BD and their affecting indicators; the combining models linked the predicted BD by each member models with the observed BD. Especially, Bayesian regularization back-propagation artificial neural network (BRANN), and adaptive neuro fuzzy inference systems (ANFIS) were employed as both member and combining models. The candidate combining models also included the simple average methods (SAM). The input variables for member models were selected by a hybrid filter and wrapper method. The performances of these models were examined using the leave-one-out cross validation. As the largest unregulated river in Alberta, Canada with ice jams frequently occurring in the vicinity of Fort McMurray, the Athabasca River at Fort McMurray was selected as the study area. The breakup dates and candidate affecting indicators in 1980–2015 were collected. The results showed that, the BRANN member models generally outperformed the ANFIS member models in terms of better performances and simpler structures. The difference between the R and MI rankings of inputs in the optimal member models may imply that the linear correlation based filter method would be feasible to generate a range of candidate inputs for further screening through other wrapper or embedded IVS methods. The SAM and BRANN combining models generally outperformed all member models. The optimal SAM combining model combined two BRANN member models and improved upon them in terms of average squared errors by 14.6% and 18.1% respectively. In this study, for the first time, the stacking ensemble learning was applied to forecasting of river ice breakup dates, which appeared promising for other river ice forecasting problems.

中文翻译:

年度河冰破裂日期的堆叠集成学习框架

摘要 河流破冰日期 (BDs) 不仅是气候变率和变化的代理指标,而且是当地冰引起的洪水管理中的一个直接问题。开发了用于年度河冰BDs的堆叠集成学习框架,其中包括两级组件:成员模型和组合模型。成员模型描述了BD与其影响指标的关系;组合模型将每个成员模型的预测 BD 与观察到的 BD 联系起来。特别是贝叶斯正则化反向传播人工神经网络(BRANN)和自适应神经模糊推理系统(ANFIS)被用作成员模型和组合模型。候选组合模型还包括简单平均方法(SAM)。成员模型的输入变量是通过混合过滤器和包装器方法选择的。使用留一法交叉验证检查这些模型的性能。作为加拿大阿尔伯塔省最大的不受管制的河流,麦克默里堡附近经常发生冰塞,因此选择麦克默里堡的阿萨巴斯卡河作为研究区域。收集了1980-2015年的分手日期和候选影响指标。结果表明,BRANN成员模型在更好的性能和更简单的结构方面普遍优于ANFIS成员模型。最佳成员模型中输入的 R 和 MI 排名之间的差异可能意味着基于线性相关的过滤方法可以生成一系列候选输入,以便通过其他包装器或嵌入式 IVS 方法进一步筛选。SAM 和 BRANN 组合模型通常优于所有成员模型。最优 SAM 组合模型结合了两个 BRANN 成员模型,并在平均平方误差方面分别提高了 14.6% 和 18.1%。在这项研究中,首次将叠加集成学习应用于河冰破裂日期的预测,这对于其他河冰预测问题似乎很有希望。
更新日期:2018-06-01
down
wechat
bug