Abstract:

To improve the operational safety and reliability of wind farms and to prevent potential safety risks, a wind power anomaly detection model based on multimodal feature fusion was proposed. First, a long short-term memory (LSTM) network was employed to extract dependency features from time-series data, by which the temporal evolution characteristics of wind power were represented. Second, the wavelet transform was adopted to convert the time series into time-frequency images, and a convolutional residual network was utilized to extract multi-scale time-frequency features. Then, a multimodal attention-based feature fusion module was designed, in which temporal features and time-frequency features were adaptively weighted and reconstructed in a unified feature space. Finally, a joint discriminative model was constructed by fusing multimodal features, and accurate detection of wind power anomalies was achieved. Experimental results show that the proposed multimodal anomaly detection model achieves significantly higher detection accuracy than traditional unimodal methods on a real-world wind farm dataset, and effectively reduces missed detection and false alarm rates. Stronger robustness is also demonstrated under complex operating conditions and sudden abnormal states.

Key words: multimodal fusion, feature fusion, wind power, anomaly detection, safety early warning, time-frequency images