The gas turbine power generation system is a core component of thermal power equipment. However, factors such as fouling on the compressor blades can lead to performance degradation. Therefore, real-time assessment of compressor performance degradation and timely implementation of water washing measures are crucial for ensuring the safe and reliable operation of the compressor. However, due to considerations of safety and economic efficiency, power plants typically do not allow compressors to operate under severe degradation conditions. As a result, acquiring data on severely degraded compressor states is challenging, making it difficult to establish traditional data-driven degradation assessment models. This limitation hinders the ability to predict the level of degradation and determine whether water washing measures are necessary. Furthermore, given that compressor performance varies under different turbine loads, this variability introduces interference when assessing the actual degradation level of the compressor, which increases the difficulty of accurately evaluating its performance degradation. To address these issues, a degradation knowledge-guided differential generative adversarial network for compressors is proposed. In the zero-shot scenario where data for severe degradation is entirely absent, expert-annotated prior knowledge is used to apply targeted deterioration to features of mildly degraded data, thereby generating features of severely degraded data. These generated features are then used to train a supervised degradation assessment model. To mitigate the interference caused by load variations, different load conditions are treated as multiple domains, and features that eliminate the impact of varying loads are extracted from operational data. Additionally, a knowledge predictor is designed to retain prior knowledge information in these features, enhancing the generation quality of the differential generative adversarial network. The effectiveness of the proposed method is validated using real-world compressor operational data. Compared to other zero-shot learning methods, the proposed method achieves a 5.22% improvement in the harmonic mean accuracy of degradation level assessment.