The implementation of deep learning-based fault diagnosis methodologies has been increasingly observed across diverse sectors within the power industry. This is particularly relevant in contexts where power stations generate vast quantities of operational data that necessitate advanced real-time processing capabilities. However, it has been observed that the efficacy of isolated deep learning models often falls below expectations, primarily due to their limited generalization capabilities. Leading to their limited application in the fault detection of power station automatic control loops. In light of these challenges, the present study introduces an innovative anomaly detection methodology specifically designed for control loops. Based on deep clustering and transfer learning. Initially, the methodology employs an autoencoder clustering algorithm to systematically categorize the operational conditions prevalent in control loops. Subsequently, the VAE-LSTM (Variational Auto Encoder-Long Short Term Memory) model is deployed to meticulously extract the latent features of the difference sequences between controlled parameters and set values. The source domain model undergoes training through the minimization of the loss function, thereby optimizing its parameters for enhanced performance. Finally, by harnessing the principles of transfer learning, the model undergoes fine-tuning of its network parameters by training the feature distribution distance of the LSTM network with target domain data, This process significantly enhances the efficiency and accuracy of fault diagnosis. Experimental validation has confirmed the method's superior performance across datasets from various domains. Moreover, it has demonstrated the capability to conduct real-time fault diagnosis across a multitude of automatic control loops.