Generalizing Time Aggregation to Out-of-Sample Data Using Minimum Bipartite Graph Matching for Power Systems Studies

This paper proposes a novel four-stage time aggregation method that can extract the underlying structure of years of historical renewable energy data. As opposed to the conventional time aggregation approaches, the proposed method starts with finding the best pair matches between patterns from one year to the next using Minimum Bipartite Graph Matching. The proposed method can capture the typical structure of patterns by taking into account the inter-annual relationships between observations collected from different years. The proposed method is evaluated by both data-based and model-based criteria. Regarding the former, we examine the quality of representative periods from different aspects such as the mean, standard deviation, maximum hourly ramp up/down, annual duration curve, and annual ramp duration curve. For the model-based evaluation, we consider a unit commitment problem in a modified IEEE 24-bus system consisting of an energy storage unit. We evaluate the methods in replicating the annual operational cost, wind energy curtailment, and energy throughput of energy storage. We evaluate the methods on three out-of-sample data sets to show the effectiveness of the proposed method in generalizing the clusters rather than over-fitting them to the in-sample observations.