Collaborative Research: Stochastic Multi-scale Optimization for Energy Resource Planning

能源投资组合的优化设计和控制需要解决一个多周期随机优化问题，该问题涵盖了数千个时间周期的细粒度和粗粒度类型的不确定性，并延伸到未来几十年。 我们需要规划对风能、太阳能、天然气、生物质能和核能等能源资源的投资，以实现特定的能源目标，除了技术、政策和气候的变化之外，还要抓住间歇性能源和需求的每小时变化。 这个问题产生了一个非常高维的随机优化问题，具有数十万个时间段。我们将结合联合收割机的近似动态规划（ADP）和机器学习的优势，以处理细粒度的不确定性来源（风能，太阳能，需求）与广义随机分解（GSD）处理粗粒度的不确定性（技术，政策和气候变化）。 使用GSD的发展将使人们有可能处理技术和政策演变中复杂的跨期依赖关系。 我们正在研究新的Dirichlet混合模型和学习率，以提高ADP算法的速度和鲁棒性，从而处理更复杂的问题。这项研究将使评估新能源发电和存储技术成为可能，通过适当考虑不确定性，并对不同技术的边际价值进行更准确的估计，比旧模型更加真实。 我们将更好地了解最重要的参数，如响应能力，存储容量和损失。 这项研究还将提高我们制定强有力政策的能力，以实现到2030年可再生能源占20%的目标。 一个更广泛的方法学的好处将是随机规划和近似动态规划领域的整合，这两个领域沿着沿着但独立的路径发展，具有明显不同的词汇表，面向不同的问题类别。

The optimal design and control of energy resource portfolios requires solving a multiperiod stochastic optimization problem that covers both fine-grained and coarse-grained types of uncertainty over thousands of time periods, extending decades into the future. We need to plan investments into energy resources such as wind, solar, natural gas, biomass and nuclear to meet specific energy goals, capturing both hourly variations in intermittent energy and demand, in addition to changes in technology, policy and climate. This problem produces a very high-dimensional stochastic optimization problem with hundreds of thousands of time periods. We will combine the strengths of approximate dynamic programming (ADP) and machine learning to handle the fine-grained sources of uncertainty (wind, solar, demand) with generalized stochastic decomposition (GSD) to handle coarse-grained uncertainties (changes in technology, policy and climate). Developments using GSD will make it possible to handle complex intertemporal dependencies in the evolution of technology and policy. We are investigating new Dirichlet mixture models and learning rates to enhance the speed and robustness of ADP algorithms to handle more complex problems.This research will make it possible to evaluate new energy generation and storage technologies with far more realism than older models by properly accounting for uncertainties and producing a more accurate estimate of the marginal value of different technologies. We will gain a better understanding of the most important parameters such as responsiveness, storage capacity and losses. This research will also enhance our ability to develop robust policies to meet goals such as 20 percent renewable by 2030. A broader methodological benefit will be the integration of the fields of stochastic programming and approximate dynamic programming, which have evolved along parallel but separate paths with distinctly different vocabularies, oriented toward different problem classes.