基于数据驱动和深度置信网络的配电网无功优化

随着分布式电源和随机负荷电动汽车等的大量接入,配电网的运行环境日益复杂,对在线无功优化及其快速性提出了更高的要求。该文将“深度学习”引入配电网无功优化,提出了基于深度置信网络的无功优化方法。通过构造高维随机矩阵,从配电网运行数据中提取统计特征作为输入,将历史控制策略进行编码作为输出,利用先无监督后有监督的方式训练深层架构,学习系统特征与无功优化策略之间的映射关系,建立基于数据驱动和深度置信网络的配电网无功优化模型。基于改造的IEEE-37节点主动配电网仿真模型,对比分析了历史数据量和分布式电源渗透率场景对传统优化方法,场景匹配方法和所提方法的无功优化效果的影响。结果表明,所提方法可明显降低网络损耗和节点电压偏移,它不依赖于系统的模型和参数,在线决策速度快,且对历史数据量要求较低,在高渗透率分布式发电等未知场景下仍能表现优良,验证了该方法的正确性、有效性和较强的鲁棒性。

Reactive Power Optimization of Distribution Network Based on Data Driven and Deep Belief Network

With a large number of distributed generators(DG) and electrical vehicles integrated into distribution system, operation complexity of the distribution system increased, arising higher requirements to fast online reactive power optimization. In this paper, "deep learning" is introduced, and an online reactive power optimization method based on deep belief network(DBN) is proposed. Firstly, utilizing the electrical and ambient big data from the distribution system operation, high-dimension random matrixes are constructed, 57 statistic features are extracted and used as the inputs of DBN, and conventional reactive power control strategies are used as the outputs of DBN. Secondly, deep architecture is trained with unsupervised and supervised methods to learn the nonlinear mapping between the system features and the reactive power strategies, and an online reactive power optimization model is established based on data driven and DBN. Based on the modified simulation model of IEEE-37 active distribution network, the effects of historical data and scenarios of distributed generation penetration on reactive power optimization results of traditional method, scenario matching method and the proposed method are compared and analyzed. The results show that, the proposed method reduces network losses and node voltage deviations obviously, not depending on distribution system model and parameter, requires less historical data, and possesses excellent performance even under unknown high DG penetration scenes, verifying correctness, effectiveness and robustness of the proposed method.