Automatically Construct Hard Unit Commitment Datasets via Data-Driven Adjustment of Demand Curve

Hard unit commitment (UC) problems require a long computing time to obtain the optimal solution, which is crucial for the development of UC algorithms. This article proposes a data-driven approach to automatically generate a set of hard UC instances that adapt to different industrial power systems. First, this article perturbs the demand curves of the original UC case to construct the initial dataset. For small-scale UC problems, by training a generative model with a loss function indicating computing time, a set of hard UC instances is generated. In addition, the generated hard instances are selected to train a new generative model. This iterative training strategy enables the generative model to focus more on latent features, progressively increasing the hardness of the generated UC instances. Second, to shorten the time required to generate hard UC instances for large-scale UC problems, we present a feature transfer method. The core idea is to use data-driven methods to extract key physical features of demand curves based on small-scale test systems that affect the UC hardness while remaining effective for large-scale systems. These key physical features are then used to constrain the construction of the initial dataset, avoiding the attempts at incorrect features during model training, which effectively reduces the time consumption of the hard case generation process. Experimental results on public test systems and industrial utility systems show that the proposed method can find the test cases with an average increase in solution time of 66.76 times without manual intervention