System Flexibility and AI Computational Enhancement for Day-Ahead Power System Operations

Smart grids play a critical part in today’s world and is paramount to optimize the usage of energy and to address the increasing penetration of renewable energy sources (RES). The power system is a complicated network of electrical elements and requires efficient operation. The day-ahead operations use the security-constrained unit-commitment (SCUC) to provide a reliable, secure, and least-cost solution while clearing the market for forecasted demand. However, existing power system operations do not use available system flexibility in the form of transmission network or demand response exhaustively. Operators rely on experience to use these resources and disregard the economic benefit of such technologies. Hence, the importance of strategies such as corrective network reconfiguration (CNR) and corrective demand response (CDR) as an economic tool are initially explored. Network reconfiguration is considered for superior economic incentive. In addition, it enables integration of RES, efficient utilization of energy storage, and reducing carbon emission in high penetration systems. This is implemented by reducing curtailment of RES and relieving system congestion, while also addressing reducing carbon emissions. However, due to the complexity added to existing SCUC model, such solutions are not scalable to practical systems. To address computational efficiency to SCUC, considering substantial economic incentive tools like CNR, two novel remedies are identified in this thesis: (1) A purely optimization-based technique is shown by utilizing benders decomposition by breaking a large SCUC model into master problem and sub-problems. The proposed approach is iteratively solved by effectively screening non-critical sub-problems to handle the computational complexity. Simulation results points to scalability to large practical power system networks. (2) A novel approach by leveraging machine learning (ML) to learn patterns between system demand profile and generator commitment schedule using historical information is developed. The ML would assist with innovative post-processing methods and create a feasibility layer to improve predictions that would result in a reduced model for problem size reduction of SCUC. The proposed approach with selective utilizing of ML predictions can bring substantial computational benefits. This is achieved without loss in solution quality while being easily extendible to any decomposed, heuristic, or sped-up algorithms for SCUC.