Abstract: The identification of dominant instability modes (DIMs) during combined transient voltage and power angle instability is critical for implementing timely power system control measures. Current data-driven approaches demonstrate limited adaptability to network topology variations, often resulting in reduced identification accuracy or complete failure. This paper introduces a topology-adaptive, data-driven DIM identification method that combines K-means clustering with multi-random convolutional kernel transformation to establish a high-precision base identification model. The K-means algorithm adaptively selects critical transient curves while the convolutional transformation effectively extracts essential DIM features including curve slope and instability duration from time-series data, ensuring robustness against topological changes. To enhance result reliability, an ensemble learning framework integrates Bagging algorithms with error-divergence decomposition theory, mitigating individual model uncertainty through optimal combination of multiple base models. Extensive testing on both the China Electric Power Research Institute's 36-node test system with modified configurations and a practical 8897-node power grid confirms the method's consistent accuracy across varying network topologies, demonstrating its effectiveness for real-world power system stability analysis.