Internal Compensation and Payment Allocation for Virtual Power Plants Under Multi-Stage Data Tampering Attacks

Virtual power plants (VPPs) use participant-reported capacity before response delivery to set internal compensation standards and participant-level response data afterward to allocate payments. Manipulation of these chronologically distinct data streams can therefore distort both physical response decisions and settlement outcomes. This paper proposes a data-tampering-aware internal compensation (DTAIC) framework that aligns protection with this information sequence. A constraint-informed conditional Wasserstein generative adversarial network with gradient penalty (CI-WGAN-GP) generates feasible temporal manipulation sequences. A two-stage adversarial soft actor-critic (SAC) policy determines cluster-specific compensation standards, while a trusted-aggregate-based joint estimator reconstructs participant responses for payment allocation. On held-out Low Carbon London data, no violations of the predefined temporal-envelope constraints were observed among 15,000 sequences generated by CI-WGAN-GP, while the proposed policy increased mean reward by approximately 10% relative to input-only adversarial SAC under combined-stage manipulation. Under severe mixed manipulation with realistic information uncertainty, joint estimation reduced participant-response error and payment transfer by roughly three quarters relative to direct allocation. The framework thus links feasible stress-scenario generation, stage-aware compensation, and aggregate-consistent payment allocation within one operational sequence.