Day-Ahead Price Scenario Generation using Conditioned Multivariate Elliptical Copulas

Forecasting day-ahead (DA) electricity market prices has become a complex challenge in recent years due to the increasingly varying daily price patterns and longer-term shocks. Traditional models using linear regression models with normally distributed error terms insufficiently model the skewed scenario distributions. On the other hand, modern nonlinear machine learning models require larger training sets, which are increasingly difficult to obtain due to the rapidly changing market dynamics. This work therefore proposes conditioned multivariate elliptical copulas for scenario generation of hourly DA market prices. We demonstrate that the t-copula captures these correlations for the Dutch DA market prices properly. To further test the quality of the generated scenario sets, their effect on a stochastic programming model is evaluated. Our use case is the maximization of the profitability of a battery system operating on the Dutch DA market. This way, we first study the stability of the operator's profit under multiple generated scenario set sizes. Then, we investigate which exogenous variables are most effective for conditioning the t-copulas. The results show that almost all models tested outperformed the naïive benchmark, with solar irradiation being the exogenous variable with the highest positive impact. The highest performance was reached with irradiation and forecasted load variables, in combination with splitting weekday and weekend models, achieving 82% of the optimal profit under perfect foresight.