부산대학교 도서관

Real-time electricity prices are economic signals incentivizing market players to support real-time system balancing. These price signals typically switch between low- and high-price regimes depending on whether the power system is in surplus or shortage of generation, which is hard to capture. In this context, we propose a new Transformer-based model to assist the short-term trading strategies of market players. The proposed model offers high-performance probabilistic forecasts of real-time prices while providing insights into its inner decision-making process. Transformers rely on attention mechanisms solely computed via feed-forward networks to explicitly learn temporal patterns, which allows them to capture complex dependencies such as regime switching. Here, we augment Transformers with subnetworks dedicated to endogenously quantify the relative importance of each input feature. Hence, the resulting forecaster intrinsically provides the temporal attribution of each input feature, which foster trust and transparency for subsequent decision makers. Our case study on real-world market data of the Belgian power system demonstrates the ability of the proposed model to outperform state-of-the-art forecasting techniques, while shedding light on its most important drivers.