Privacy-Preserving Load-Shaping Strategies for Smart Meters using Deep Reinforcement Learning

Smart meters are important for improving households’ energy management and efficiency, but they raise significant privacy concerns by potentially revealing detailed user behaviors through energy consumption patterns. Recent work have proposed methods using physical energy resources to modify smart meter readings; however, they neglect the importance
of keeping the energy storage device’s electricity level consistent after the protection cycle. In this paper, we present PLS-DDPG, a novel proactive load-shaping strategy based on
the deep deterministic policy gradient (DDPG) algorithm. The proposed method enhances privacy by creating artificial load patterns and masking original load signatures by incorporating a novel reward-shaping mechanism to incentivize privacy-centric
behaviors. We also present a battery consistency mechanism to ensure efficient energy management. The proposed method is evaluated on a state-of-the-art adversary algorithm: TCN-based non-intrusive load monitoring (NILM). We trained two NILM algorithms for kettle and toaster usage disaggregation because these two appliances are commonly used and can reflect user life cycles. The performance of our proposed method is compared with other baseline algorithms. The evaluation results indicate that our proposed algorithm is on par with the previous DQN-based work in privacy protection, with both demonstrating
top-tier performance among all baselines evaluated. However, PLS-DDPG retains 96.55% of the electricity capacity compared to 67.24% for the DQN-based work, highlighting its superior
efficiency in managing energy resources while upholding privacy.