From Short-Term Signals to Long-Term Trends: Efficient Forecasting of Balancing Market Price Trends

Sammanfattning

Electricity price forecasting has been widely studied in short-term contexts, particularly within day-ahead markets, where data-driven models have demonstrated strong performance. Conversely, long-term forecasting of balancing market prices remains relatively underexplored, with existing approaches primarily relying on complex, data-intensive structural models. This highlights a research gap regarding efficient, long-term forecasting approaches. Following that gap, the research question of this study is: “Which data-driven models, originally developed for short-term electricity price forecasting, are most suitable for efficient forecasting of long-term balancing market price trends?” To answer the research question in the context of the Swedish market, a set of explanatory variables representing structural characteristics of the power system are utilized to forecast the trend component of balancing price premiums, extracted using Empirical Mode Decomposition, with a 5 year horizon. A Random Forest model and a Multilayer Perceptron are trained under a time-based data split, evaluated using accuracy- and efficiency-based metrics, and compared against benchmark models from the literature. The results indicate that the Multilayer Perceptron is the most suitable model among those evaluated, achieving comparatively stronger predictive performance. The findings suggest that structural variables carry predictive information for long-term balancing price dynamics, and that neural network architectures are more transferable to long-term forecasting contexts than linear or tree-based approaches. However, no model demonstrates reliably strong performance across contexts, and results are subject to important limitations regarding data availability and distributional shifts. These findings suggest that data-driven approaches leveraging publicly available structural variables may offer a viable and efficient alternative to structural simulation models for long-term balancing market forecasting