Data-Driven Dynamic Probabilistic Reserve Sizing Based on Dynamic Bayesian Belief Networks

Due to the variability and partial unpredictability of renewable energy sources (RES), generation reserve sizing is key for managing operational risks and enhancing grid reliability. Most existing approaches for reserve sizing estimate the forecasting error for RES using either simple point forecasts, or probabilistic forecasts that assume a predefined parametric model of the underlying distribution, such as a Gaussian distribution. However, these approaches are unable to accurately model the probability of critical events that occur in the tails of the distributions, which is vital for operational risk management. Further, they do not account for multiple dynamic factors that may impact operational risk. In this paper, we introduce a data-driven dynamic probabilistic reserve sizing method based on the artificial intelligence technique of dynamic Bayesian belief networks. The method considers the actual underlying distribution of forecasting errors from the availability of conventional generators (forecasting errors of their available generation capacity), demand, RES, and prevailing conditions such as weather and market prices. In addition, we introduce a new dynamic metric for calculating the reliability level of the power grid in order to provide a real-time stochastic decision support tool for power system operators. The proposed method is demonstrated using seven years of real historical data with a granularity of 30 min from the power system in Australia, and has been implemented and is currently used by the Australian system operator.