潮流计算是电力系统的基本问题之一，也是对系统进行稳定分析和最优潮流计算等的基础。虽然已有的牛顿类方法可以在理论上较好地解决潮流计算问题，但是在调整不收敛潮流初值等问题上仍然存在困难。本文利用在图像识别等领域取得较大成功的深度学习模型，探究将其应用于电力系统潮流初值判敛及不收敛潮流初值调整问题的可行性。本文首先建立了以多隐藏层人工神经网络（DNN）和卷积神经网络（CNN）为基础的潮流初值判敛模型，进一步结合生成式对抗神经网络（GAN），构造了不收敛潮流初值的调整模型，并且在所生成的训练集和测试集上对所提算法的有效性进行了验证。论文的主要工作如下：（1）从潮流初值判敛问题出发，较为详细地介绍了如何基于最基础的深度学习算法-DNN搭建深度学习网络、基于IEEE14节点系统介绍了构造训练集和测试集的步骤、提出了用于训练模型的特征向量的构造方法，并且给出了基于可视化方法和统计方法的误判样本分析思路；（2）由于传统CNN无法解决在非欧几里得数据上构造卷积核的问题，首先利用权重矩阵介绍了基于空间位置的图卷积思路；然后，运用图拉普拉斯变换，给出了将谱域图卷积方法应用至电力系统潮流初值判敛问题的思路；最终，利用IEEE39节点系统生成的数据集进行了算例验证；（3）通过调整GAN中判别器与生成器的损失函数构造了潮流初值调整模型，该模型以不收敛的潮流初值作为输入，通过模型中的生成器给出了潮流初值的调整方向；接着，通过在判别器与生成器的损失函数中增加罚项，使该潮流调整模型在训练过程中纳入了电力系统潮流调整的实际约束条件；最终，利用IEEE14以及IEEE39节点系统构造了训练集与测试集，验证了所提潮流调整模型的有效性。综上，本文从牛顿类潮流计算方法出发，基于深度学习网络得到了可以应用于潮流初值判敛问题的深度学习模型，并且进一步构造了调整不收敛潮流初值的深度学习模型，给出了一种可行的潮流初值调整方案，为实际电网进行潮流初值调整提供了一种新的思路。

Power flow calculation is a basic problem in power system, and is also the basis of voltage stability analysis, transient stability analysis and optimal power flow calculation. Although Newton’s method can solve power flow calculation problem theoretically, it still encounters difficulties when dealing with modulation of non-convergent power flow. This thesis takes advantages of deep learning models, which have achieved great success in the area of image recognition, and explores the feasibility of using these models to solve problems in power flow calculation. This thesis first builds models based on DNN and CNN to solve the convergence criterion of power flow calculation and models based on GAN to provide feasible modulation solution for non-convergent power flow, and then validates the models by samples of train sets and test sets generated in this thesis. The main contributions of this thesis are presented as follows: (1) Based on the process of convergence criterion of power flow, this part introduces the process of modelling of DNN, creation of train sets & test sets based on IEEE14 system, selection of feature vectors based on the topology of power system & crucial information for power flow calculation, and visualization method & analyzing method of misjudged samples;(2) Since traditional CNN cannot be applied to data of non-Euclidean structure due to the convolutional kernel problem, this part first provides a feasible solution to build convolution kernel based on the weight matrix. Then according to Graph Laplace transformation, this part introduces the graph convolution method in the spectral demain to the convergence criterion of power flow and validates the method based on the samples of train sets and test sets generated from IEEE39 system;(3) By changing the loss functions of discriminator and generator models of GAN, this part proposes a feasible modulation solution for non-convergent power flow, for systems with and without constraints, and finally validates the method based on the samples of train sets and test sets generated from IEEE14 and IEEE39 systems.This thesis proposes deep learning models based on DNN and CNN which can be applied to convergence criterion of power flow calculation and models based on GAN which can be applied to modulation solution for non-convergent power flow. This work is a feability study of applying deep learning to the core calculation of power system, and also an enlightening method for modulation of non-convegent power flow of real power systems.