Learning�Assisted Constraint Filtering to Enhance Power System Optimization Performance

Abstract

Machine learning (ML) is a powerful tool that provides meaningful insights for operators to make fast and efficient decisions by analyzing data from power systems. ML techniques have great potential to assist in solving optimization problems within a shorter time frame and with less computational burden. AC optimal power flow (ACOPF), dynamic economic dispatch (D-ED), and security-constrained unit commitment (SCUC) are the three energy management optimization functions studied in this dissertation. ACOPF is solved every 5~15 minutes. Because of the nonconvex and complex nature of ACOPF, solving this problem for large systems is computationally expensive and time-consuming. Classification and regression learning models are used to identify inactive transmission line flow constraints and omit them from the optimization, thus relieving computational costs. D-ED is solved daily and hourly to determine the best generation schedule. A combined learning and model-based algorithm is developed to identify the status of network and thermal unit ramp-up/down constraints and formulate a reduced-size dynamic economic dispatch problem. The learners read network topology, demand, and thermal unit generation cost information as input and identify the status of network and ramp constraints as output. SCUC solved daily is a complex decision-making problem that belongs to the class of mixed-integer optimization category. Benders decomposition is an effective method for solving this class of problems but suffers from exponential worst-case computational complexity. Classification and regression learners are used to enhance the convergence performance of Benders decomposition. The learning models generate tighter cuts and filter out non-useful cuts at each Benders iteration to accelerate a two-stage stochastic SCUC problem.