Real-Time Neural Distributed Energy Resources Dispatch with Feasibility Guarantees

As renewable energy sources like solar and wind become more prevalent, grid operators need high-frequency real-time scheduling to handle their variability. Neural network-based surrogates offer fast computation, but they struggle to strictly enforce nonconvex power flow constraints without resorting to slow external solvers. This paper from Zhu, Xu, and Sun bridges that gap with a novel solver-free neural dispatch framework that guarantees feasibility. The key innovation is a convex inner approximation of the DistFlow model (the standard power flow model for distribution networks) derived via the convex envelope theorem. Building on that, the authors formulate a robust optimization-based affine policy that yields a theoretically certified interior-point mapping rule.

This mapping rule is then embedded in a bisection-based projection scheme that efficiently corrects infeasible neural network outputs—all without any external solver. Experimental results show the method restores feasibility on the order of 10^{-3} seconds while maintaining near-optimal performance. For energy system operators, this means they can rely on fast neural dispatchers for real-time control without sacrificing safety or optimality, enabling more aggressive renewable integration. The work is particularly relevant for distribution system operators managing behind-the-meter solar, battery storage, and other distributed energy resources.