From Linear Regression to Temporal Fusion Transformers: Benchmarking Electricity Price Forecasting Models in Markets with Different Renewable Energy Shares

Abstract

Rising renewable penetration increases wholesale price volatility and challenges traditional forecasting approaches. This paper benchmarks three methods—adaptive linear regression, Random Forest (RF), and the Temporal Fusion Transformer (TFT)—on hourly day-ahead prices in four European markets with contrasting generation mixes: Slovakia (nuclear-dominated), Spain (solar-rich), Denmark (wind-dominated), and Germany (mixed). Using Energy-Charts data for 2015–2025, all the models are evaluated under one day-ahead protocol on a common 12-month hold-out against naive 24-h persistence. With price-history features, the RF significantly outperforms persistence in three of four markets. A single cross-market five-seed TFT ensemble attains the lowest MAE everywhere, cutting the persistence error by 55.6% in Germany (to 12.88 €/MWh) and 21.7% in Slovakia. The improvement is generally larger in markets with higher variable-renewable penetration, though not strictly monotonic. Replacing the realized fundamentals with TSO day-ahead forecasts changes the MAE by at most 0.22 €/MWh: the day-ahead price is itself formed from those forecasts. Multivariate Diebold–Mariano tests confirm both advantages everywhere on absolute errors. Quantile averaging yields pinball losses of 3.87–6.53 €/MWh, but the 80% intervals cover only 71–78%. Variable importance ties the leading driver to each mix: wind in Denmark and Germany, load and solar in Spain, and load in Slovakia.