Transformer-based forecasting with synthetic input data generation for day-ahead electricity markets

Abstract

Our research paper proposes a unified forecasting framework applicable across all pan-European markets that are part of the European Network of Transmission System Operators for Electricity (ENTSOE) association covering 36 countries across Europe, providing a standardized approach to electricity price forecast (EPF) for day-ahead markets (DAM). The framework is based on a hybrid approach that combines Transformer Architecture (TA), synthetic input and inverse optimization. The framework addresses an important challenge in EPF: the unavailability of day-ahead predictions for renewable (RES) generation (photovoltaic, wind) and sold, which are highly correlated with price volatility. To overcome this limitation, synthetic inputs are generated to obtain new features, such as sold and RES contribution to load. These synthetic features are generated using a multi-objective optimization method that minimizes forecasting errors, aligns synthetic inputs with historical patterns and ensures temporal smoothness. The forecasting framework is developed using an extensive feature engineering process, incorporating weather predictions, load forecasts and advanced volatility indicators, such as rolling statistics, exponential moving averages and Bollinger Bands. To assess the framework’s performance and generalization across pan-European markets, simulations were done for Romania, Spain, Poland, Finland and the Czech Republic. The framework performed with an R2 between 0.95–0.98 and Mean Average Error (MAE) between 2.24–15.01 EUR/MWh on the test set, and R2 between 0.91–0.97 and MAE between 5.39–19.58 EUR/MWh on the evaluation set.