부산대학교 도서관

Unit commitment (UC) is a computationally challenging problem with the increasing modeling scale and tight time limit. As a mixed-integer linear programming (MILP) problem, the unique problem-oriented properties need to be utilized to accelerate the UC calculation. In this paper, we propose to extract the key information of the UC solution using a small-scale adjoint model and use the information to guide the acceleration of UC without any loss of accuracy. To formulate the adjoint model, we first fix a majority of the integer variables based on a linear programming (LP) relaxation. Then, a feasibility repair strategy is proposed to obtain a feasible solution of the original UC. Finally, we use a machine learning-based clustering approach to avoid unsuitable fix of integer variables, aiming to improve the solution quality of the adjoint model. With the solution quickly obtained from the adjoint model, we can get necessary information of the UC problem ahead of the formal solving process. Further, the adjoint model is utilized in two aspects: set the lazy constraints according to root node solution and exchange incumbents to get better upper bounds in the branch & bound (B&B) process. The whole procedure does not make any modifications to the original UC model and hence has the optimality guarantee. Case studies show that the proposed method can achieve an acceleration of UC calculation from 1.64 to 4.74 times in 30 test cases of public RTS-GMLC system and practical utility system in China.