<p>As a vital component of modern critical infrastructure, the secure operation of smart grids is of paramount importance. With the expansion of smart grid scale and the increasing complexity of network attack methods, traditional intrusion detection methods face significant challenges in terms of accuracy and adaptability. To address these challenges, this paper proposes a deep ensemble learning-based intrusion detection system (DEL-IDS) for smart grids, which accurately identifies various intrusion behaviors and enhances the reliability of intrusion detection. Specifically, a deep ensemble learning architecture based on stacking is designed. This architecture fully leverages the temporal feature extraction capabilities of RNN, GRU, and BiLSTM, combined with self-attention mechanisms and a stacked strategy, to effectively capture long-term dependencies in the data and significantly improve the model’s ability to recognize complex attack patterns. Furthermore, a novel metaheuristic algorithm, the elite chaotic particle swarm optimization (ECPSO), is proposed to effectively address the hyperparameter optimization problem in smart grid intrusion detection. The chaotic mapping initialization strategy introduced in ECPSO enhances population diversity, while the elite retention mechanism significantly improves the global search capability and convergence efficiency of ECPSO. To validate the effectiveness of the DEL-IDS model, comparative experiments are conducted on multiple datasets. The results demonstrate that the model outperforms advanced methods such as Adaboost, BiLSTM, Spider, DIS-IOT, and CNN-LSTM in terms of accuracy, precision, recall, and F1 score, effectively enhancing the accuracy of smart grid intrusion detection. Specifically, the DEL-IDS model achieves multi-class classification accuracies of 95.82%, 97.67%, and 99.72% and multi-class classification F1 scores of 95.03%, 97.66%, and 91.34% on the MSU-ORNL, CIC-IDS2017, NSL-KDD datasets, respectively.</p>