In the power domain, the feedback validation of model accuracy for large-scale power-specific models constructed based on professional expertise is a critical technical approach to enhance the model’s integration effectiveness with newly added data samples. Establishing a data sample perturbation evaluation system and methodology based on the fusion of knowledge graphs and large-model feedback is an inevitable choice to mitigate the adverse effects of anomalous data samples. To this end, this paper proposes an anomalous data mitigation method incorporating the Conditional Generative Adversarial Network (CGAN) algorithm to impose conditional constraints on the knowledge generation process. It defines a control generator based on knowledge graph constraint information to extract the original anomalous key features of generated samples and adjusts the dataset’s anomaly coefficient based on multi-round constraint feedback. To reduce the involvement of noise vectors, physical constraints, statistical constraints, and semantic constraint boundaries are extracted from the knowledge graph information. A multi-task discriminative structure is employed to simultaneously evaluate sample authenticity, mitigation degree, and physical constraint satisfaction. Leveraging a joint cyclic feedback mechanism between the knowledge graph and the large model, the method inversely maps the correspondence between generation results and sample descriptions, computes the semantic similarity between original and mitigated samples, preserves critical semantics through contrastive learning, and compresses out-of-bound features to diminish the impact of anomalous data samples. This approach conditionally reduces the feature correlation between anomalous and normal data samples to within a target range, thereby improving the business semantic quality of sample data to enhance the model’s application efficiency.