Abstract: The rapid integration of renewable energy and the continuous expansion of its installed capacity have fueled profound and fundamental transformations in the operational mechanisms of modern power systems. These developments have introduced enhanced variability and volatility into system behavior, causing operational states to become increasingly complex, nonlinear, and dynamic. Consequently, conventional operating mode identification approaches, which are predominantly based on empirical rules and static system assumptions, are inadequate for capturing the evolving patterns of new power systems, especially under high levels of uncertainty and renewable penetration. To address the increasingly prominent challenges in new power systems, this study proposes a systematic and data-driven framework for identifying and analyzing operating modes. The framework begins with a data preprocessing phase that fully accounts for the characteristics of multi-source data. Targeted strategies for outlier detection and missing-value imputation are applied based on the statistical distribution of different variables, which ensures the integrity, consistency, and reliability of the input data at the source and lays a robust foundation for subsequent modeling and analysis. To compensate for data sparsity and imbalance, which are common under high renewable penetration conditions, the framework incorporates a generative module based on an Adversarial Autoencoder that integrates Temporal Convolutional Networks and Long Short-Term Memory. Through this hybrid architecture, the model can effectively learn the latent properties of the data while generating realistic and diverse augmented samples to address the problems of data imbalance. Additionally, a cosine annealing learning-rate schedule is employed during model training to enhance learning stability, prevent convergence to local minima, and improve overall training efficiency and representational quality. To address the issue of high-dimensional data and extract essential latent representations, an Autoencoder is pre-trained to compress the operational data into a low-dimensional feature space. The resulting compact and informative features are then used as input to a Dirichlet process gaussian mixture model (DPGMM), which is employed for clustering and operation mode identification. As a nonparametric Bayesian approach, DPGMM can adaptively infer the appropriate number of clusters without requiring manual specification. Such adaptive capability significantly enhances the model's flexibility, scalability, and generalization capacity. Furthermore, the framework employs the Uniform Manifold Approximation and Projection algorithm to perform dimensionality reduction and visualize the distribution of operating modes in a three-dimensional space. This enables deeper insights into the structural evolution of the system’s operational states. The proposed framework is validated using real-world operational data from a power system in a city in North China. Experimental results demonstrate that the method exhibits excellent performance in identifying operation modes under complex and highly dynamic conditions. Specifically, as the penetration level of renewable energy increases, the number and dispersion of operating modes increase significantly. The number of modes increases from three in low-penetration scenarios to six under medium penetration and up to nine in high-penetration cases. Quantitative analysis further reveals that system dynamics evolve with stronger nonlinearity and increased uncertainty and that the randomness of transitions between modes substantially increases. These findings highlight the limitations of traditional rule-based dispatch strategies and emphasize the urgent requirement for intelligent, adaptive, and flexible control mechanisms to ensure a safe, stable, and efficient operation of future power systems.