A new approach to subseasonal multi-model forecasting: Online prediction with expert advice
A new approach to subseasonal multi-model forecasting: Online prediction with expert advice
复制标题
次季节多模型预测的新方法：根据专家建议进行在线预测
DOI：
10.5194/egusphere-egu2020-17663
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
F. Ziel
中科院分区：
文献类型：
--
作者：
D. Brayshaw;P. Gonzalez;F. Ziel
<div>
<p>多模式组合在气候预测中的好处以前已经介绍过，并描述了不同的时间尺度（例如，Siebert and斯蒂芬森2019，DelSole 2007，Sansom et al. 2013）。最典型的组合方法涉及加权策略，为每个模型分配一个恒定的因子，或者统一分配，或者通过技能评估分配。考虑到模型的技能可能在不同的时间尺度上变化，并且由于多种原因（例如，季节性变化的技能，或由于预测系统的变化），这些权重保持不变的事实引入了限制。一百六十; </p>
</div><div>
<p>在机器学习领域，已经开发了一系列算法来执行专家建议的在线预测（Cesa-Bianchi et al. 2006）。这些方法考虑一组加权专家（通常在过程开始时均匀加权）来产生随后的预测，其中组合或“混合”被更新以优化损失或技能函数。一百六十; </p>
</div><div>
<p>这些在线预测方法在气候预测中的应用可能具有以下几个优点： </p>
</div><div>
<div>
<ul><li>在每个预报步骤中更新专家组合的事实允许系统在某些条件下进行调整（例如，以上所述，以保持技能。 </li>
<li>由于不同的组合可以很容易地获得不同的分位数的predictand分布，一个强大的系统可以训练，最大限度地提高技能，其全范围。一百六十;</li>
<li>由于混合模型能够适应并抛弃不称职或起反作用的专家（或赋予他们最小的权重），因此将这些专家的风险降到最低。一百六十;</li>
</ul></div>
<div>
<p>这些在线预测方法的另一个潜在应用可能是在亚季节到季节意义上的无缝预测系统的设计，这是S2 S4 E（https：//s2s4e.eu/）等几个研究项目所感兴趣的。例如，可以用一组专家来训练系统，该组专家包括子季节预报的后续启动以及季节预报的先前启动。如果在任何时候有有用的信息产生于较长的提前时间的季节性预测，混合物将赋予它更高的权重。一百六十;</p>
</div>
<div>
<p>一套这些在线预测方法已在S2 S4 E项目中进行了测试，并与更典型的多模型组合技术进行了比较，以评估其对预测国家一级能源需求和潜在其他变量的有用性。结果表明，这些创新的方法表现出显着的技能提高（高于5%），相对于更标准的技术和个人预测系统的提前4周。一百六十; </p>
</div>
</div>
<div>
<p>The benefits of multi-model combinations in climate forecasting have been previously introduced and described for different temporal scales (e.g., Siebert and Stephenson 2019, DelSole 2007, Sansom et al. 2013). Most typical combination methodologies involve weighting strategies that assign each model a constant factor, either uniformly or through a skill assessment. Given that the skill of the models can vary at different timescales, and for multiple reasons (for example, seasonally varying skill, or due to changes in the forecasting system), the fact that these weights remain constant introduces limitations.  </p>
</div><div>
<p>Within the realm of Machine Learning, a family of algorithms have been developed to perform ‘online prediction with expert advice’ (Cesa-Bianchi et al. 2006). These methods consider a set of weighted ‘experts’ (usually uniformly weighted at the start of the process) to produce subsequent predictions in which the combination or `mixture’ is updated to optimize a loss or skill function.  </p>
</div><div>
<p>These online forecasting methods potentially have several advantages for their use in climate prediction: </p>
</div><div>
<div>
<ul><li>The fact that the expert combination is updated in every forecast step allows the system to adjust in certain conditions (e.g., the ones mentioned above) to preserve skill; </li>
<li>Since a different combination can be easily obtained for different quantiles of the predictand distribution, a robust system can be trained that maximizes skill for its full range. </li>
<li>The risk of including incompetent or counterproductive experts is minimized by the fact that the mixture is able to adapt and discard them (or assign them minimal weights). </li>
</ul></div>
<div>
<p>Another potential application of these online prediction methods could be on the design of ‘seamless’ forecasting systems in the sub-seasonal to seasonal sense, which is of interest to several research projects such as S2S4E (https://s2s4e.eu/). For example, the system could be trained with a set of experts that include subsequent launches of a sub-seasonal forecast as well as prior launches of a seasonal forecast. If at any point there is useful information arising from the longer lead time seasonal forecast, the mixture would assign higher weights to it. </p>
</div>
<div>
<p>A set of these online prediction methods have been tested within the S2S4E project and compared to more typical multi-model combination techniques to assess their usefulness for the prediction of country-level energy demand, and potentially other variables. Results show that these innovative methods exhibit significant skill improvements (higher than 5%) with respect to more standard techniques and to individual forecasting systems for lead weeks up to 4.  </p>
</div>
</div>