Unit commitment and economic dispatch underpin secure and economical grid operation, yet classical mixed-integer programming explodes in complexity as networks grow, while vanilla deep learning ignores hard physics, yielding infeasible schedules. We present an end-to-end framework that couples a Graph Attention Temporal Convolutional Network with a Physics-Informed Constraint Calibrator. To address the mixed-integer complexity, the network learns spatio-temporal dispatch patterns, whereas the calibrator strictly enforces generator capacities, ramp rates, minimum startup/shutdown time logic, power balance and AC network transmission capacity limits through a decoupled discrete-continuous optimization layer, obviating the need for external solvers. Evaluated on IEEE 30- and 118-bus test systems benchmarked with high-resolution real-world load profiles from Southwestern China, the framework demonstrates superior robustness, maintaining a strict zero violation rate even under extreme/unseen scenarios and measurement errors. Extensive comparative experiments show that our method delivers schedules with near-optimal secondary costs, reaching a negligible optimality gap of +0.0149% vs. Gurobi while achieving inference speeds approximately 30 × faster than hybrid solver-based strategies. These results confirm that integrating explicit physical reasoning via a deterministic projection and physics-regularized end-to-end mechanism) within neural architectures provides a license-free, practical path to real-time, reliable power system scheduling, closing the gap between predictive accuracy and enforceable feasibility. • A hybrid GNN framework with a decoupled discrete-continuous physics layer. • Iterative PICC strictly enforces the minimum start–stop times and the constraints in continuous dispatch. • Achieves near-optimal cost (+0.0157% gap) versus commercial Gurobi/CPLEX. • Maintains 0% constraint violation rate even under extreme/unseen scenarios and measurement errors. • Accelerates scheduling by orders of magnitude via license-free tensor operations.
Xu et al. (Tue,) studied this question.