Deep learning-based rolling horizon unit commitment under hybrid uncertainties

Unit commitment is an optimization problem in power systems, which aims to satisfy future load at minimal cost by scheduling the on/off state and output of generation resources like thermal units. One challenge herein is the uncertainties that exist in both supply and demand sides of power systems, which becomes more severe with the growing penetration of renewable energy and the popularity of diversified loads. This paper proposes a rolling horizon model for unit commitment optimization under hybrid uncertainties. First, a probabilistic forecast approach for future load and wind power is given by exploiting the advanced deep learning structures, i.e. long short-term memory neural networks. Second, a Value-at-Risk-based unit commitment model is applied to decide the on/off state and output of thermal units in the next 24 h. Then at each time window, the distributions of future load and wind power are dynamically adjusted by a rolling forecast mechanism to involve the real-time collected data, whereafter a look-ahead economic dispatch model is applied to improve the output of units. Finally, the effectiveness of this research is demonstrated by a series of experiments. Generally, this study introduces a fundamental way to integrate forecast approaches into classical unit commitment optimization models.