Heterogeneous Ensemble Learning to Solve Sub-Hourly Unit Commitment

To address the growing volatility in net load, the sub-hourly unit commitment (UC) problem has gathered wide attention and is commonly formulated as a mixed-integer linear programming (MILP) problem. To leverage historical data and accelerate the solution process, this paper proposes a data-driven solution method for sub-hourly UC. A heterogeneous ensemble learning method is established by combining k-Nearest Neighbors (kNN), Support Vector Machine (SVM), and Classification and Regression Trees (CART) classifiers to improve the prediction. The method predicts and fixes parts of binary commitment and start-up variables, and pre-processes and removes corresponding minimum up time constraints coupling them, so that a reduced MILP problem that is easier to solve is obtained. Numerical testing results demonstrate that our method significantly reduces the solution time compared to using the commercial solver alone, while the solution is very close to the optimal one.