深度强化学习算法以数据为驱动,且不依赖具体模型,能有效应对虚拟电厂运营中的复杂性问题。然而,现有算法难以严格执行操作约束,在实际系统中的应用受到限制。为了克服这一问题,提出了一种基于深度强化学习的改进深度Q网络(improved dee...深度强化学习算法以数据为驱动,且不依赖具体模型,能有效应对虚拟电厂运营中的复杂性问题。然而,现有算法难以严格执行操作约束,在实际系统中的应用受到限制。为了克服这一问题,提出了一种基于深度强化学习的改进深度Q网络(improved deep Q-network,MDQN)算法。该算法将深度神经网络表达为混合整数规划公式,以确保在动作空间内严格执行所有操作约束,从而保证了所制定的调度在实际运行中的可行性。此外,还进行了敏感性分析,以灵活地调整超参数,为算法的优化提供了更大的灵活性。最后,通过对比实验验证了MDQN算法的优越性能。该算法为应对虚拟电厂运营中的复杂性问题提供了一种有效的解决方案。展开更多
As an emerging paradigm in distributed power systems,microgrids provide promising solutions to local renewable energy generation and load demand satisfaction.However,the intermittency of renewables and temporal uncert...As an emerging paradigm in distributed power systems,microgrids provide promising solutions to local renewable energy generation and load demand satisfaction.However,the intermittency of renewables and temporal uncertainty in electrical load create great challenges to energy scheduling,especially for small-scale microgrids.Instead of deploying stochastic models to cope with such challenges,this paper presents a retroactive approach to real-time energy scheduling,which is prediction-independent and computationally efficient.Extensive case studies were conducted using 3-year-long real-life system data,and the results of simulations show that the cost difference between the proposed retroactive approach and perfect dispatch is less than 11%on average,which suggests better performance than model predictive control with the cost difference at 30%compared to the perfect dispatch.展开更多
文摘深度强化学习算法以数据为驱动,且不依赖具体模型,能有效应对虚拟电厂运营中的复杂性问题。然而,现有算法难以严格执行操作约束,在实际系统中的应用受到限制。为了克服这一问题,提出了一种基于深度强化学习的改进深度Q网络(improved deep Q-network,MDQN)算法。该算法将深度神经网络表达为混合整数规划公式,以确保在动作空间内严格执行所有操作约束,从而保证了所制定的调度在实际运行中的可行性。此外,还进行了敏感性分析,以灵活地调整超参数,为算法的优化提供了更大的灵活性。最后,通过对比实验验证了MDQN算法的优越性能。该算法为应对虚拟电厂运营中的复杂性问题提供了一种有效的解决方案。
基金partially supported by Hong Kong RGC Theme-based Research Scheme(No.T23-407/13N and No.T23-701/14N)SUSTech Faculty Startup Funding(No.Y01236135 and No.Y01236235).
文摘As an emerging paradigm in distributed power systems,microgrids provide promising solutions to local renewable energy generation and load demand satisfaction.However,the intermittency of renewables and temporal uncertainty in electrical load create great challenges to energy scheduling,especially for small-scale microgrids.Instead of deploying stochastic models to cope with such challenges,this paper presents a retroactive approach to real-time energy scheduling,which is prediction-independent and computationally efficient.Extensive case studies were conducted using 3-year-long real-life system data,and the results of simulations show that the cost difference between the proposed retroactive approach and perfect dispatch is less than 11%on average,which suggests better performance than model predictive control with the cost difference at 30%compared to the perfect dispatch.
