With the large-scale integration of renewable energy into power systems, operating scenarios have become increasingly complex and diverse, and the accurate identification of extreme operating scenarios can no longer be effectively achieved by traditional experience-based methods. To address this issue, a data-driven method for the identification of extreme operating scenarios is proposed. Firstly, an improved auxiliary classifier generative adversarial network (ACGAN) is constructed to conditionally generate extreme operating scenarios to different risk categories. Then, the generated scenarios are incorporated into the real operating scenario set, and a support vector machine (SVM) is trained using labeled data to accurately identify extreme operating scenarios. Finally, the proposed method is validated on the IEEE 39-bus system. The results show that the extreme scenarios generated by the improved ACGAN are highly consistent with the distribution of real scenarios, effectively alleviating the data imbalance caused by the scarcity of extreme operating scenarios. Moreover, the sample-enhanced SVM model can accurately identify extreme operating scenarios with different risk levels. Reliable technical support is therefore provided for decision-making related to the safe and stable operation of power systems.