A Neural Combinatorial Optimization Algorithm for Unit Commitment in AC Power Systems

The unit commitment (UC) problem in AC power systems can be formulated as a mixed-integer nonlinear optimization program with a running time that scales exponentially with the number of generators. This paper addresses the time complexity of solving the UC problem by developing a deep learning framework that determines the generator on/off states using a transformer deep neural network (DNN), and subsequently solves an AC optimal power flow (OPF) problem to obtain the generator setpoints. To obtain a feasible binary solution, we apply a neural combinatorial optimization algorithm to train the DNN, while penalizing infeasible power flow solutions. Also, to guarantee the optimality of the generator setpoints, we transform the AC OPF problem into a semidefinite program (SDP). The proposed algorithm can obtain a near-optimal solution to the UC problem in polynomial running time. Simulations are performed for two IEEE test systems. When compared with three existing UC algorithms in the literature, our proposed algorithm can obtain a solution with at least 2.14% lower operation cost and lower running time. When compared with the MOSEK solver, our algorithm can obtain a solution with at most 1.97% greater operation cost, but with a significantly lower running time.