融合多类型深度迁移学习的电力系统暂态稳定自适应评估

针对不同类型人工智能网络应用于电力系统暂态稳定评估时精度和泛化能力不稳定、运行方式或拓扑结构发生较大变化时评估精度下降、重新训练新模型费时费力等问题，提出一种融合多类型深度迁移学习模型（tmDLM）的自适应评估方法，该方法融合了深度置信网络、卷积神经网络以及长短期记忆网络3种不同的深度学习模型。将训练好的各类深度学习模型作为源域模型，当运行方式或拓扑结构发生较大变化时，采用少量目标域样本集微调预训练模型，使其快速跟踪系统当前的运行状态，并得到tmDLM。新英格兰10机39节点系统和华中电网的仿真结果表明：所提方法可以充分发挥各类深度学习方法的优势，具有良好的泛化能力；六分类模型能够在判稳的同时进行稳定裕度/失稳程度等级的评估；经过迁移后的深度学习模型具有良好的评估精度和时效性，大幅缩短了模型更新时间，实现了电力系统暂态稳定的自适应评估。

Adaptive assessment of transient stability for power system based on transfer multi-type of deep learning model

Aiming at the problems that the accuracy and generalization ability are unstable when different types of artificial intelligence networks apply in transient stability assessment of power system, the assessment accuracy reduces when the operation mode or topological structure changes greatly, and the time and effort are wasted when retraining a new model, an adaptive assessment method integrating with a transfer multi-type of deep learning model(tmDLM) is proposed, which integrates with three different deep learning models of deep belief network, convolutional neural network and long short-term memory network. Each type of the trained deep learning model is taken as the source domain model, when the operation mode or the topological structure changes largely, a small number of the object domain sample sets are used to fine-tuning the pre-trained model, making it quickly track the current operation state of the system, then the tmDLM is obtained. The simulative results of New England 10-machine 39-bus system and the Central China Power Grid show that the proposed method can make full use of the advantages of each type of deep learning method with good generalization ability, the six-classification model can assess the stability margin/instability degree while judging the stable state, the deep learning model has good assessment accuracy and timeliness after transfer learning, which greatly reduces the update time of the model and realizes self-adaptive assessment of transient stability of power system.