Can model weighting improve probabilistic projections of climate change?
Can model weighting improve probabilistic projections of climate change?
复制标题
模型加权可以改善气候变化的概率预测吗？
DOI：
10.1007/s00382-011-1217-8
复制
发表时间：
2012
期刊：
影响因子：
4.6
通讯作者：
J. Ylhäisi
中科院分区：
文献类型：
--
作者：
J. Räisänen;J. Ylhäisi
最近，Räisänen和合著者提出了一个加权方案，其中多模式集合中可观测气候和气候变化之间的关系决定了与观测的一致性在多大程度上影响气候变化预测中的模式权重。在第三次耦合模式相互比较项目（CMIP 3）数据集内，该方案略微提高了确定性温度变化预测的交叉验证精度。在这里，相同的方案适用于概率温度变化预测，在强烈的限制假设下，CMIP 3合奏跨越实际建模的不确定性。交叉验证表明，概率温度变化预测也可以通过这种加权方案得到改善。然而，相对于均匀加权的改进在尾部敏感的对数分数中比在连续排名的概率分数中小。在世界大部分地区，加权对预测真实世界的21世纪温度变化的影响不大。然而，在主要是高纬度海洋的某些地区，分布的平均值发生了很大的变化和/或分布范围大大缩小。各个模型的权重随着位置的变化而变化很大，因此在某些区域获得几乎为零权重的模型在其他地方仍然可能获得很大的权重。虽然这种变化的细节是特定的方法，它表明，不同的模型的相对优势可能难以利用加权计划，使用空间统一的模型权重。
Recently, Räisänen and co-authors proposed a weighting scheme in which the relationship between observable climate and climate change within a multi-model ensemble determines to what extent agreement with observations affects model weights in climate change projection. Within the Third Coupled Model Intercomparison Project (CMIP3) dataset, this scheme slightly improved the cross-validated accuracy of deterministic projections of temperature change. Here the same scheme is applied to probabilistic temperature change projection, under the strong limiting assumption that the CMIP3 ensemble spans the actual modeling uncertainty. Cross-validation suggests that probabilistic temperature change projections may also be improved by this weighting scheme. However, the improvement relative to uniform weighting is smaller in the tail-sensitive logarithmic score than in the continuous ranked probability score. The impact of the weighting on projection of real-world twenty-first century temperature change is modest in most parts of the world. However, in some areas mainly over the high-latitude oceans, the mean of the distribution is substantially changed and/or the distribution is considerably narrowed. The weights of individual models vary strongly with location, so that a model that receives nearly zero weight in some area may still get a large weight elsewhere. Although the details of this variation are method-specific, it suggests that the relative strengths of different models may be difficult to harness by weighting schemes that use spatially uniform model weights.