Abstract: In view of the common problems of high false alarm rate and poor adaptability to complex power consumption patterns in current power anomaly detection methods for smart hospitals, a hybrid prediction model integrating CNN, LSTM and attention mechanism is proposed, and a dynamic threshold determination method based on kernel density estimation is innovatively introduced. Extract the local and periodic features of hospital power load by using CNN; model long-term dependencies through LSTM; by leveraging the attention mechanism to focus on the critical periods of abnormal signs, the prediction accuracy and abnormal sensitivity are significantly enhanced. In actual deployment, dynamic thresholds are adopted to replace the traditional fixed thresholds to adapt to the normal drift of the hospital's power consumption baseline. The experimental results show that this method improves the accuracy of anomaly detection by approximately 15% compared to the traditional LSTM model, reduces the false alarm rate by 20%, and can effectively locate typical anomalies such as line aging and sudden equipment failures, providing reliable technical support for the power safety and refined management of smart hospitals.