<p>Wind-power curves form a critical basis for wind farm condition monitoring and anomaly type recognition. Current supervisory control and data acquisition (SCADA)-based approaches primarily rely on single-modal numerical or visual features, which constrains their ability to capture both global morphology and local structural deviations in power curves. Furthermore, the shortage of labeled abnormal samples restricts recognition performance. To tackle these challenges, this article proposes an image-topology-semantic multi-feature fusion method under a contrastive learning framework. SCADA scatter data are transformed into grayscale images, from which global image features are extracted using a Vision Transformer. Local topological structures are modeled via Holistically-Nested Edge Detection and graph convolutional networks, while textual anomaly descriptions are encoded by a CLIP text encoder and aligned via a parameterized dual-path fusion module. By jointly integrating global morphology, local topology, and textual physical semantics, the proposed framework resolves the difficulty of distinguishing visually similar wind-power curve anomaly types under limited labeled abnormal samples. The method utilizes a two-stage training strategy: pre-training on external curve datasets and fine-tuning on a simulated wind-power curve dataset, followed by evaluation on real wind farm SCADA data under both known-category anomaly-type recognition and leave-one-anomaly-type-out zero-shot recognition scenarios. Experimental results indicate that the proposed method achieves 89.5% accuracy and 89.4% F1-score with limited-data fine-tuning, 94.2% accuracy and 94.1% F1-score with sufficient-data fine-tuning, and a 78.1% F1-score in leave-one-anomaly-type-out zero-shot recognition, demonstrating its effectiveness for predefined wind-power curve anomaly type recognition in data-scarce settings.</p>