Sensitivity analysis of factors relevant to extreme imbalance between procurement plans and actual demand: Case study of the Japanese electricity market

In a liberalized power system, operators control the power system to compensate for imbalances, which are the differences between the scheduled power procurement in electricity markets and the actual power supply. Interpreting the influence of factors that affect extremely large imbalances (i.e., extreme imbalance events) enables power system operators to implement appropriate system operation plans. Conventionally, such imbalance analysis has focused only on a limited number of factors and described variation in imbalance by utilizing highly interpretable statistical models, assuming that the probability of imbalance events varies monotonically with changes in those factors. In particular, sensitivity analysis of such models can be a powerful tool to support the decision-making of system operators. However, when dealing with the increasingly complex behavior of markets involving many actors, a flexible statistical analysis framework is required to identify informative factors among the various observable quantities, and to describe nonmonotonic relationships when essentially necessary. This study focuses on the statistical behavior of the odds ratio of the extreme imbalance events by concentrating on an inherently large number of explanatory variables. The authors propose a model-based approach using a class of partially linear additive models and a variable selection method to reveal the statistical relationships between the relevant variables and extreme imbalance events. The framework further provides an analysis scheme to determine the response sensitivity of relevant variables to the odds ratio of extreme imbalance events. The usefulness of the framework was demonstrated by applying the approach to a real-world dataset collected in the Japanese electricity market system. The results show that the proposed approach based on partially linear additive models works well to describe the extreme imbalance events; the constructed models derive the interpretable sensitivity curves to clarify the impact of informative variables to extreme events, while identifying monotonicity/nonmonotonicity among variables.