基于深度学习的直流配电网分布鲁棒优化调度方法

随着分布式资源的大规模接入，直流配电网能量损耗小、控制灵活的优点凸显。针对直流配电网传统物理优化模型效率低的问题，提出了一种基于深度学习的直流配电网分布鲁棒优化(DRO)调度方法，其采用深度学习方法替代了基于场景的DRO模型的迭代求解过程，通过直接预测典型场景的最恶劣概率分布来提高模型求解效率。构建直流配电网基于场景的DRO物理模型，采用列与约束生成算法迭代求解生成深度学习的训练数据；以光伏出力、负荷、范数置信度为输入，以最恶劣概率分布为输出，构建深度神经网络模型；基于训练好的神经网络预测实时输入的光伏出力、负荷、范数置信度的最恶劣概率分布，构建最恶劣概率分布下的单层随机规划模型，获取等效的基于场景的DRO调度策略；采用33节点直流配电网系统为算例，验证所提方法在求解效率和计算精度方面的有效性。

Distributionally robust optimal dispatching method of DC distribution network based on deep learning

With the large-scale access of distributed resources, the DC distribution network advantages of small energy loss and flexible control are prominent. A distributionally robust optimal(DRO) dispatching method of DC distribution network based on deep learning is proposed for low efficiency of the traditional physical optimization model for DC distribution network, which adopts the deep learning method to replace the iterative solving process of the scenario-based DRO model. By directly predicting the worst probability distribution of typical scenario, the efficiency of the model solution can be improved. The scenario-based DRO physical model of DC distribution network is constructed. The column and constraint generation algorithm is used for iterative solution to generate the training data of deep learning. Then, the deep neutral network model is built with photovoltaic output, load power and norm confidence are taken as the input and the worst probability distribution is taken as the output. Based on the trained neural network, the worst probability distribution of the real-time input of photovoltaic output, load power and norm confidence is predicted. The single-layer stochastic programming model under the worst probability distribution is constructed to obtain the equivalent scenario-based DRO dispatching strategy. The 33-bus DC distribution network system is used as an example to verify the effectiveness of the proposed method in terms of solution efficiency and calculation accuracy.