基于深度迁移学习的时变拓扑下电力系统状态估计

State Estimation for Power Systems with Time-varying Topology Based on Deep Transfer Learning

针对电力系统拓扑实时变化导致数据驱动状态估计器不可用的情况,提出一种基于深度迁移学习的数据驱动状态估计方法。将原拓扑海量历史数据训练得到的模型作为基础模型,当新拓扑实时量测数据更新时,加载和保存基础模型中特征提取层的权重和参数,只需要微调模型的全连接层,即可获得适应于新拓扑的神经网络,提高了数据驱动状态估计模型的自适应性和泛化性能。通过对IEEE标准系统和中国某实际省网的算例测试,并将其估计结果与加权最小二乘法和加权最小绝对值法进行比较。结果表明,在考虑拓扑时变性的情况下,该算法与上述2种物理算法相比具有更优的估计性能和估计效率。

A data-driven state estimation method based on deep transfer learning is proposed for the situation that the data-driven state estimator is not available due to the real-time change of power system topology.The model obtained by training the massive historical data of the original topology is used as the base model.When the new topology is updated with real-time measurement data,the weights and parameters of the feature extraction layer in the base model are loaded and saved.Only fully connected layers of the model need to be fine-tuned to obtain a neural network adapted to the new topology,which improves the adaptiveness and generalization performance of the data-driven state estimation model.The estimation results are tested by arithmetic cases on the IEEE standard system and an actual provincial power grid of China,and compared with the weighted least squares method and weighted least absolute value method.The results show that the algorithm has better estimation performance and estimation efficiency compared with the above two physical algorithms when topological time-variability is considered.