Forecasts of imbalance prices are essential for participants in short-term electricity markets. This study proposes an approach to integrate trade data from continuous intraday markets into forecasting methods and considers the production of both parametric and non-parametric density forecasts. Extensive feature engineering is required to convert irregular trade data into a suitable tabular structure. Our findings demonstrate that intraday trade data can enhance state-of-the-art forecasts based on quantile regression and LASSO. We also propose a model based on Natural Gradient Boosting to predict the means and variances of a bimodal Gaussian mixture, which offers both improved predictive performance and computational efficiency. In a case study based on multiple years of data from Great Britain, our proposed approach reduced the Pinball Loss of state-of-the-art models by 20% for forecasts made 30 seconds before gate closure, and by 7.5% for forecasts made one hour before delivery.
🌊
