The power industry faces a fundamental computational trade-off: DC optimal power flow models run quickly and scale to large networks, but their solutions often violate the nonlinear AC power-flow equations that govern actual grid physics. System operators have traditionally accepted this gap or spent significant computation time converting DC solutions to AC-feasible points. A new parameterized restoration framework addresses this challenge through machine learning-optimized mapping.
The PACR method uses differentiable surrogates to implement two key correction mechanisms. First, distributed slack variables balance active power imbalances across multiple buses rather than concentrating correction at a single location. Second, PV/PQ bus switching adjusts how generators regulate reactive power as voltage conditions change. Rather than using fixed, one-size-fits-all parameters, the approach trains multiple tunable parameters—including slack participation factors, voltage setpoints, and regulation steepness—offline using the implicit function theorem and automatic differentiation.
Once parameters are optimized during a training phase, they remain fixed and enable rapid AC restoration during real-time operations. This maintains the computational advantage of DC-based dispatch while substantially improving feasibility and cost outcomes.
Testing on IEEE, ACTIVSg, and PEGASE benchmark systems demonstrates significant improvements over conventional single-slack recovery methods. On the largest case with 9,241 buses, the optimized approach reduced cost differences by 80% compared to baseline AC recovery and solved 75% faster than full AC optimal power flow. Performance gains persist across systems of varying sizes and configurations.
The technique bridges a persistent gap in power system operations: the need for both computational speed and physical constraint satisfaction. By leveraging modern differentiation tools to optimize restoration parameters, the method provides operators with a practical solution that maintains DC's scalability while substantially improving solution quality. This could enable faster, more reliable dispatch optimization in markets and control centers worldwide.
