Emotion recognition from Electroencephalogram (EEG) signals has emerged as a promising method for understanding human affective states. However, Deep learning-based emotion recognition models suffer from overfitting and generalisation due to the variability in EEG signals and the scarcity of labelled data, which impede their performance. In this work, a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) based architecture was adopted for efficient EEG data augmentation. The publicly available “EEG Brainwave” dataset was used to train the WGAN-GP model to synthetically generate the fake EEG data. The generated synthetic data was mixed with the real data in different proportions to determine the optimum ratio of data augmentation for efficient emotion classification. The efficacy of the data augmentation was evaluated by proposing an LSTM-based classifier that efficiently classifies the three emotional states: positive, neutral, and negative. The experimental results show that the maximum classification accuracy of \(99.14\%\) was achieved with a precision of 0.9915, recall of 0.9914, and F1 score of 0.9914 when an equal quantity of real and synthetically generated EEG data was mixed to train the classifier. Our WGAN-GP-LSTM method not only enhances the robustness of emotion recognition models by utilizing data augmentation but also significantly improves the classification accuracy with limited labelled data and outperforms all other state-of-the-art techniques.