In the context of the rapidly evolving cyber-physical landscape of modern power systems, ensuring robust and real-time network security at the edge has become a critical challenge. This paper proposes a novel lightweight and embedded network behavior recognition framework based on Data Processing Units (DPU), designed specifically for power communication networks. The proposed method integrates multi-scale traffic decomposition via Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), adaptive sub-sequence generation using Partial Autocorrelation Function (PACF), and complexity-aware temporal modeling through PeepHole Long Short-Term Memory (LSTM) networks. To address the issue of imbalanced datasets, a Generative Adversarial Network (GAN) module is introduced to synthesize minority attack traffic samples, thereby enhancing the model’s generalization capability. Based on the results on the NSL-KDD and Bot-IoT datasets, the proposed model outperforms the compared models in terms of accuracy, training time, and model size.