11 September 2025 in IEEE Transactions on Smart Grid
The large-scale adoption of smart meters has facilitated the collection of numerous amounts of high-resolution load values from consumers, supporting a range of load consumption studies that contribute to enhancing the overall efficiency of smart grid. However, the analysis of true load data may potentially reveal private information, which users may be reluctant to share. Data perturbation or data synthesis mechanisms are typically employed to address the issue, whereas these methods are either inflexible (i.e., difficult to freely adjust privacy intensity) or still lack theoretical guarantees to alleviate privacy breaches (i.e., unable to effectively prevent inference attacks). The generative model trained with differential privacy can effectively mitigate the above challenges. However, there exists a series of limitations such as complex training procedures and the risk of mode collapse in previous work on privacy-preserving training, which significantly hinders the utility of load data generation. To this end, a privacy-preserving load profile synthesis framework PrivLoad that combines the diffusion model and differential privacy is proposed in this work, allowing to retain the statistical features of the true load while safeguarding user privacy. In particular, a denoising mechanism considering the range of load values is devised to capture the temporal correlations of real load profiles and enhance the fidelity of synthetic loads. Comprehensive case studies on the real load dataset are performed to assess the performance of proposed framework. The experimental results from various aspects reveal the superiority of PrivLoad. For instance, when the privacy budget is 10, PrivLoad achieves an improvement of more than 40% in clustering metrics compared to the baseline. Moreover, the real-world application of peak load estimation highlights the practicality of the designed approach, which can achieve an outstanding performance that is extremely close to the original load dataset.
28 May 2024 in IEEE Transactions on Smart Grid
The growing development of the fast charging of electric vehicles has strengthened the interdependence between urban transportation networks (UTN) and power distribution networks (PDN). Through cooperative price guidance of UTN operator and PDN operator, the performance of the above networks can be improved coordinately. However, vehicle travelers may not want to share their actual traveling and charging demand information with the UTN and PDN operators for pricing due to privacy concerns. In order to protect travelers privacy while reducing the system operation cost, this paper proposes a privacy-preserving cooperative pricing framework PrivCPM for the power-traffic networks. In particular, a novel privacy preservation mechanism, DPTC, based on the Differential Privacy technique for Traffic and Charging demands is designed with a physics-constrained post-processing (PP) process, which is developed to enhance the utility of the perturbed data significantly (i.e., improve the feasibility of the cooperative pricing optimization problem and reduce the system operation costs) by integrating the actual network constraints and the correlation between various types of vehicles. Comprehensive case studies are conducted based on a 12-node UTN and a 20-node PDN. The results show that, compared to the operation cost obtained using the true data, PrivCPM under the privacy budget (i.e., the upper limit of the amount of information that can be disclosed) of 0.01 can obtain the near-optimal solution with only approximately 2.6% higher cost. Moreover, the proposed PP process can significantly enlarge the range of travelers privacy budget, and the operation cost obtained by using PrivCPM can achieve a 3.35% cost reduction but with a five times lower privacy budget compared to the conventional method (i.e., PrivCPM without PP). In addition, PrivCPM in the above setting merely increased the cost by 7.31% compared to the non-private case.
1 November 2026 in IEEE Transactions on Knowledge and Data Engineering