融合关键节点信息的数据驱动暂态电压稳定性评估模型优化方法OA

A new approach for power system transient voltage stability assessment(TVSA)is offered by deep learning models,yet challenges persist in two key aspects:the interpretability of assessment results(e.g.,how critical node characteristics are identified by the model)and the controllability of the decision-making process(e.g.,how assessment logic for specific instability modes is adjusted).The mechanism by which the model focuses on critical nodes struggles to be revealed by traditional methods;moreover,when assessment rules for specific instability modes are optimized,the assessment logic of other modes tends to be disrupted by them,resulting in insufficient model reliability in complex power grid scenarios. To improve the interpretability and controllability of TVSA models,an optimized training method integrating critical node information is proposed in this paper,which consists of three core steps:First,a TVSA model based on a Transformer encoder is constructed.Spatiotemporal features of bus node voltage phase angle sequences are learned by this model via a self-attention mechanism,and the model's critical node assessment rules are visualized using attention weights.Second,the fault-induced delayed voltage recovery index(FDVRI)is introduced.Based on the dynamic recovery characteristics of node voltages within 0～4 cycles after fault clearance,critical node sets for each training sample are labeled to clarify the core feature carrier of the instability mode.Third,a loss function incorporating an attention guidance term and an attention retention term is designed.Critical nodes of the target instability mode are forced to be focused on by the guidance term through weight constraints,thereby correcting the model's assessment rules;the stability of attention distribution for non-target modes is maintained by the retention term,avoiding interference of the optimization process on global assessment logic. Validations are conducted on the IEEE 39-node system and the Northwest China Power Grid.Results show that the model's attention mechanism toward critical nodes is clearly revealed by attention visualization,and the influence of critical nodes on assessment decisions is directly reflected by their attention weights.Dominant instability features are accurately mined by the model during training through the loss function integrating critical node information:while the assessment ability for specific instability modes is optimized,the physical consistency of the model's overall assessment logic is ensured.After a specific instability mode is optimized using the proposed method,the accuracy of samples for this mode is increased by 8.72%,with no impact on the assessment of other instability samples. The interpretability and controllability of TVSA models are enhanced by this optimized method through the introduction of critical node labeling and a dual-constraint loss function,providing an intuitive basis for operators to understand assessment logic.It is confirmed by validations on both standard systems and actual power grids that instability mode features are effectively captured by the method,offering an interpretable and controllable new path for power system TVSA.