深度学习驱动的微网群互动合作与优化运行研究

With the increasing national attention and promotion on “Internet +” Intelligent Energy, interconnected multi-microgrid with multi-energy incorporation and open interaction constitutes a promising direction for future energy internet construction. The coordination of regional multi-microgrid cooperation and interaction with the utility grid by regional energy management companies or retail companies with multiple investments and participants is one of the most important development modes for microgrid cluster. However, the conventional model-based multi-microgrid cooperative gaming and coordinated optimization problem is afflicted by parameter uncertainty, model complexity, and information incompleteness. It’s due to the above considerations that this project proposes a deep-learning based cooperative interaction framework for microgrid cluster within the grand architecture of energy internet, which involves the modeling of interactive response among microgrid cluster, the computation of Nash Equilibrium, as well as sequential decision-making optimization of coordinated operation by applying deep-learning technique. Further endeavor would be devoted to the establishment of a deep-learning based multi-microgird interactive simulation and decision-making optimization platform with the combination of the existing hybrid MATLAB/GAMS optimization platform, which will realize the interactive comparison and test analysis of different decision-making schemes, and provide a convincing theoretical and technical foundation for a rationalized, orderly, and large-scale development of microgrid cluster with multiple investments and multiple participants within the framework of energy internet in near future.

随着国家对“互联网+”智慧能源的关注和推进，以多源融合与开放互动为主要特征的互联微网群将成为未来能源互联网建设的重要方向，“多方投资、多方参与”的区域能源管理公司或是售电公司统一协调区域内多微网合作互动，并与电网协同运行将会是微网群重要发展模式之一。但基于模型优化的多微网合作互动博弈和协同优化问题求解存在难以应对不确定性、模型需要简化、信息需要完全透明等障碍。本项目即是在此背景下，提出基于深度学习驱动的微网群合作互动机制和框架的新思路，解决微网群互动响应特征的深度学习建模、合作互动Nash均衡的强化学习求解、深度学习与模型结合的协同运行时序决策等一系列基础问题。结合已有的基于MATLAB/GAMS微网群互动仿真平台，建立基于深度学习驱动的微网群互动仿真和决策优化平台，实现不同方案的交互对比分析与测试，为“多方投资、多方参与”的微网群合理、有序地规模化发展及互动提供理论基础和技术支撑。

多微网合作互动并与电网协同运行是微网群重要发展模式之一。本项目针对基于模型优化的多微网合作互动博弈和协同优化问题求解方案仍存在难以应对不确定性、模型需要简化、信息需要完全透明等障碍，研究提出了基于深度学习驱动的微网群合作互动机制和框架的新思路，从微网群互动响应特征的深度学习建模、合作互动Nash均衡的求解、深度学习与模型结合的协同运行时序决策等三方面开展了研究。首先，针对微网群互动响应特征分析中所面临的信息不完备、特性差异大，难以精确建模等问题，提出了不完备信息下基于LSTM长短时记忆深度网络的微网群互动行为等值建模方法，证明了提出LSTM深度学习的微网群互动行为建模方法的精确性和可行性。然后，以风-光-氢构建的多主体微网为对象，建立了风–光–氢多主体合作运行纳什谈判模型，并将其等效转化为联盟效益最大化和电能交易支付谈判两个子问题，采用交替方向乘子法进行分布式求解，解决了风电、光伏和电制氢多类主体微网的合作优化运行及收益公平分配两大关键问题。最后针对微网群之间的实时互动高效决策问题，提出了基于微网深度学习封装建模的微网群数据驱动互动定价方法，在无需掌握各微网内部参数数据的情形下，直接基于各微网互动的外部环境封装模型进行奖励函数和受益的估值，进而结合数据驱动优化方法的高维空间搜索及寻优决策的能力，实现更加高效的微网群互动运行。针对结合已有的微网群互动仿真平台，搭建了基于深度学习驱动的微网群互动仿真和决策优化平台，实现了不同方案的交互对比分析与测试，本项目成果可为“多方投资、多方参与”的微网群合理、有序地规模化发展及互动提供理论基础和技术支撑。