Photovoltaic (PV) power simulation is indispensable for the planning and operation of renewable-rich power systems. Among various simulation methods, data-driven indirect simulation methods are the most efficient, primarily involving scenario extraction and power output generation. Typical scenario extraction methods often feature coarse characterizations, which also cannot be expressed by traditional power output generation models lacking the ability of quantifying the uncertainty of the output results. This paper proposes a fine-grained simulation model for PV power output interval, i.e., upper bound and lower bound, based on two-stage clustering of multiple typical scenarios and dual-ensemble compatible learning. The first stage of scenario extraction, considering the impact of total daily irradiance on PV output, employs a line distance index for preliminary scenario division of daily PV output. These scenarios are further divided into five characteristic scenarios with an angular distance index considering the impact of intra-day irradiance variations on power output. PV power output interval is generated from dual-ensemble compatible learning including Stacking and Bagging, which are adaptively selected with a Bias-Variance Selection Index (BS) calculated using the value of scenario characteristics. Stacking and Bagging are driven by three neural network base learners with Bayesian layers, which can output the lower bound and upper bound with confidence at a certain level, thus quantifying the uncertainty of the result accurately. The actual data of Daxing and Gangjialing power stations are used in case study. Results form case study show that compared with the existing methods, the proposed method can improve the Power Coverage Rate (PCR) and Average Power Simulation Accuracy (APA) of uncertainty simulation results by 27.44% and 5.39%, respectively. Meanwhile, the Average Daily Maximum Power Simulation Accuracy (AMPA) is similar to that of existing methods.
Read full abstract