Smart meters (SM) have generated vast amounts of detailed load data, enabling advanced load profile analyses that can significantly enhance smart grid efficiency. However, this data collection raises serious privacy concerns, as it can inadvertently expose sensitive user information. Traditional methods to address these concerns, such as data perturbation or synthetic data generation, often lack flexibility or fail to fully mitigate privacy risks. To address these challenges, we propose a Transformerbased conditional Generative Adversarial Network designed to generate differentially private (DP) synthetic SM data from the original meter data. Our approach produces high-quality DP-synthetic SM load data that inherently satisfies user-level differential privacy, ensuring secure data analysis. Extensive experiments demonstrate that our method not only serves as a robust alternative to the original dataset for real-world load forecasting but also effectively preserves user anonymity. Additionally, we demonstrate that service providers can optimize energy consumption by predicting future occupancy status while maintaining user privacy with minimal error, using previous occupancy data and DP-synthetic load data. Our experimental results show that the synthetic SM data generated by our method is approximately 35% more efficient than the synthetic data produced by the conditional Transformer-GAN proposed in prior research. Moreover, our DP-synthetic data offers around 82% greater privacy protection compared to previously established DP-only privacy-preserving techniques. Finally, our evaluation of the model’s resilience against adversarial attacks reveals a 28% reduction in load forecasting performance degradation when using our DP-synthetic data, compared to models trained on the original SM data.