As China advances its dual carbon strategy, integrating new energy sources into power grids has grown significantly, making power system operations more complex and dynamic. For deep learning-based models used in transient stability assessment to be reliable, the training data and the data encountered in real-world applications must be independent and identically distributed. However, because power systems are time-varying and uncertain, models trained offline may not perform well in new operational scenarios. This paper proposes a transient stability assessment-discriminative domain adaptive (TSA-DDA) framework to address variations in operating scenarios.

Firstly, an inter-domain dual distribution adaptation method was proposed. While aligning the marginal probability distributions of the source and target domains, this method also used Bayes' theorem to align the conditional probability distributions, achieving optimal domain adaptation. Secondly, both mean and variance differences between the source and target domains were comprehensively considered in the domain adaptation process. A new transfer regularization term was constructed to measure the inter-domain distribution differences, improving the model's domain adaptation capability. Finally, a discriminant Softmax function with adjustable parameters was developed to make intra-class sample features more compact while keeping inter-class sample features away by adjusting the parameters. This improvement can enhance the applicability of the assessment model to power grids.

In the case studies, the TSA-DDA framework's ability to address variations in operational scenarios was first validated on the New England 10-machine 39-bus system. Subsequently, four alternative TSA-DDA frameworks, each with specific modules removed, were established to evaluate the effectiveness of individual components. The prediction accuracy of the target and source domain test sets was compared using a fine-tuning algorithm and the TSA-DDA. The TSA-DDA’s capacity for continual learning is confirmed. The TSA-DDA was then benchmarked against mainstream transferred learning approaches to verify its effectiveness in scenarios with limited new data. Finally, to assess the generalization capability of the proposed scheme, experiments were conducted on a larger and more complex provincial power grid in Southwest China. The experimental simulations utilized the PSD Power Tools and Dynamic Simulation Program to offer high-fidelity power system simulation data for model training and testing.

The conclusions of this paper are given as follows. (1) The inter-domain dual distribution adaptation method comprehensively measures differences in marginal and conditional probability distributions between domains from both mean and variance perspectives. It constantly forces the feature extractor to narrow these differences, ensuring effective feature alignment across domains and enhancing the model’s adaptability. (2) The discriminant Softmax function improves the model’s learning of discriminative features by compacting intra-class features and separating inter-class features, which enhances the performance of the domain adaptation framework in transient stability assessment tasks. (3) Using voltage trajectory clusters with clustering and convergence properties as model inputs, the proposed framework ensures effective transferability across systems with varying structures and scales.