Stanford researchers slash power grid AI training time 4x

Researchers from Stanford University's Department of Electrical Engineering have developed DiffAPQP, a solver-flexible framework that dramatically accelerates AI training for power system optimization. The open-source Python package (available on GitHub) tackles a major bottleneck in decision-focused learning (DfL) for power networks by eliminating the need to repeatedly solve complex optimization problems during training.

The innovation comes from two key technical advances: first, automatic canonicalization of quadratic power-system models written in CVXPY into differentiation-ready representations, and second, exploiting repetitive solving structures through solver warm-start and data updates. For the backward pass, the team established mathematical equivalence between differentiation through full KKT systems and reduced systems, enabling envelope-theorem-based gradients that avoid expensive adjoint KKT solves. On the IEEE 118-bus system with 24-hour coupled economic dispatch and redispatch, DiffAPQP achieved 2.27x-3.58x speedups in closed-loop training and 3.62x-4.38x speedups in counterfactual training compared to CvxpyLayers, with best configurations reaching 3.91x and 6.39x improvements respectively.