<p>Emotion recognition based on Electroencephalogram (EEG) plays a significant role in the field of human–computer interaction and neurofeedback applications. Although Deep Learning (DL) methods possess excellent capabilities in automatic feature extraction, developing techniques that efficiently extract discriminative features from EEG signals is still a challenge in dealing with the emotion recognition task. Hence, this research proposes the Prairie Wolf Optimization-C2A-Attention model (PWO-C2A-AM model) for emotion recognition using EEG signals. Specifically, the proposed model exploits the Prairie wolf optimization for selecting the informative electrodes that capture the biological topology information of the brain region resulting in improving the recognition accuracy. Simultaneously, the C2A attention mechanism adaptively selects the most important regions and suppresses the irrelevant features leading to extracting the discriminative spatial–temporal features from EEG channels. Furthermore, the proposed PWO-C2A-AM model incorporating an attention mechanism processes the learned representations and prioritizes the most relevant time steps resulting in accurate emotion recognition. In addition, the PWO algorithm optimally tunes the weights and biases of the PWO-C2A-AM model resulting in improving the accuracy and reducing the error. Extensive experiments demonstrate that the proposed PWO-C2A-AM model attains high accuracy, sensitivity, and specificity of 98.09%, 99.27%, and 96.91% during the 90% of training for the DEAP dataset. Further, the proposed model obtains superior results achieving the high accuracy, sensitivity, and specificity of 98.31%, 97.19%, and 99.43% for 90% of training using the SEED dataset. Besides, the proposed approach outperformed other existing techniques by achieving the high accuracy, sensitivity, and specificity of 97.06%, 98.21%, and 95.84% for 90% of training using the PME4 dataset.</p>