Content area
The stochastic unit commitment problem seeks to determine the day-ahead generating unit on/of schedule in a power system with significant renewable generation. In this context, this dissertation leverages data-driven approaches to balance solution efficiency and modeling accuracy. First, since solving the stochastic unit commitment problem is computationally very costly, we propose two practical data-driven approaches to reduce its computational burden. Second, we build a computationally efficient security-constrained stochastic unit commitment model to identify accurate reserves by considering uncertainties of renewable generation and demand, and generating unit failures. Third, we design two risk metrics to rank committed units based on the impact their failures have on system operation.
Since the time available to solve the stochastic unit commitment problem is limited, the first research work in this thesis aims to shorten the solution time of the stochastic unit commitment problem by reducing its size via learning. In particular, learning from historical solved instances, we fix unchanged binary variables and eliminate inactive inequality constraints in the stochastic unit commitment problem. The numerical results show that the reduced problem generally requires significantly less time to solve while obtaining high-quality solutions, which are very close to or indistinguishable from those obtained by solving the original problem.
The second research work in this thesis seeks to drastically reduce the computational burden of the network-constrained stochastic unit commitment problem by using a quasi-deterministic proxy. The proxy has the same structure as the original stochastic unit commitment problem but only includes two envelope scenarios, generated based on a large scenario set representing renewable generation uncertainty and demand uncertainty. The proxy can identify very similar commitment decisions as the one obtained by solving the original problem. Its computational performance, though, is close to that of a deterministic unit commitment problem.
The third research work in this thesis develops a computational tractable security-constrained stochastic unit commitment model to determine the optimal reserve level. The proposed model characterizes the uncertainty of renewable generation and demand via scenarios and enforces security constraints per scenario to ensure an uneventful operation in case of failure of key thermal units. Once a contingency occurs, the reserve obtained by the proposed model can be effectively deployed to prevent unserved energy.
The fourth research work in this thesis designs two metrics to rank committed units according to the risk their failures bring to the system operation. Considering an extreme scenario, the first metric provides multiple ranking criteria. The second metric ranks units based on the system flexibility reduction once the considered unit fails. Once units are committed, those two metrics make the system operator aware of next-day operation risks.