Benchmark-aware short-term electricity price forecasting and estimation in Spain: information-set analysis, explainability, uncertainty, and renewable sensitivity

Electricity price prediction in renewable-rich markets requires not only accurate models, but also clear benchmark design, interpretable results, realistic information-set definitions, and uncertainty assessment. This study develops a benchmark-aware short-term electricity price forecasting and estimation framework for the Spanish market using a public hourly energy–weather dataset. Three main benchmark settings are considered: (A) the operator-issued day-ahead price as a practical reference signal, (B) machine-learning models without forecast-oriented inputs, and (C) forecast-aware machine-learning models that additionally include day-ahead price, load-forecast, solar-forecast, and wind-forecast information. Under the current aligned setting, CatBoost was the strongest model family and achieved a MAE of 1.405 €/MWh, a RMSE of 1.910 €/MWh, and a R 2 of 0.972 on the held-out test set. A stricter ex ante setting is also evaluated, in which contemporaneous system-state and weather aggregates that are not guaranteed to be available at the forecast origin are removed; in that setting the main performance conclusion remains largely unchanged, with only a small deterioration in test accuracy. Additional lag-only and lag-plus-calendar benchmarks show that price memory forms the predictive core of the problem, but that the full model still provides statistically significant incremental gains, especially in high-renewable, high-volatility, peak-hour, and upper-tail conditions. Explainability, ablation, bootstrap, Diebold–Mariano, and interval-calibration analyses further support the interpretation of the framework. Overall, the contribution of the study lies not in a new forecasting algorithm, but in an integrated and information-set-explicit empirical evaluation framework for short-term electricity price prediction in Spain.