Forecasting Imbalance Price Densities With Statistical Methods and Neural Networks

Abstract

Despite the extensive research on electricity price forecasting, forecasting imbalance prices is a relatively new topic. Interest, however, is growing because of the greater uncertainties and costs involved in real-time balancing. Whilst there has been previous work on nonlinear statistical methods, this article reports on a comparative study involving these and various neural network architectures including N-BEATS, fully connected, attention-based, and recurrent neural networks. To ensure valid comparability, these different neural networks were tested on the same data from Britain used in the previous point and density forecasting research. While there are only marginal improvements in point forecasts, we find that neural networks produce significantly more accurate density forecasts. Since the risks involved with exposure to imbalance prices are becoming a serious consideration for market participants, accurate density forecasts are crucial for risk management.