Computationally efficient data synthesis for AC-OPF: Integrating Physics-Informed Neural Network solvers and active learning

This study addresses the challenges of privacy, utility, and efficiency in releasing privacy-preserving operational data for AC Optimal Power Flow (AC-OPF) research. Traditional methods, injecting noise into operational data (i.e., demand data and dispatch profiles) within the Differential Privacy (DP) framework, often violate physical constraints within the data, leading to unrealistic and infeasible outcomes that diminish data utility. While AC-OPF-solver-based bi-level post-processing optimizations can enforce physical feasibility, the objective divergence between post-processing and AC-OPF leads to discrepancies, compromising data utility. Additionally, their non-convex and adversarial nature makes computation prohibitively expensive, further preventing efficient data release. To overcome these challenges, our research introduces a DP approach that combines strategic noise injection for demand data with the computation of corresponding dispatch profiles, ensuring the privacy-preserving data satisfy AC-OPF’s physical constraints. To accelerate data release, we employ Physics-Informed Neural Networks (PINNs). This ensures solutions’ physical feasibility while enhancing computational efficiency. Furthermore, we incorporate active learning to target the most informative data samples, enhancing PINN training and optimizing efficiency while maintaining solution accuracy. Comprehensive experiments on IEEE test systems reveal our approach’s improved performance and accelerated computation speed over traditional methods, highlighting its efficiency in maintaining data privacy and utility and decreasing computational burden amidst diverse privacy considerations.