From 1 to 48 hours: preprocessing strategies and machine learning forecasting for solar, photovoltaic and wind power in the Spanish intra-day electricity market

This paper presents a methodology for hourly forecasting of solar PV, solar thermal, and wind power generation in the Spanish intraday market using advanced Machine Learning techniques. Historical generation data (2018–June 2024) from Spain's Market Operator (OM) and meteorological data from a grid covering the country were utilized. A key challenge was the traceability of generation facilities, complicating data preprocessing for some technologies. The preprocessing phase included selecting critical exogenous variables like solar irradiance, wind speed, and temperature using statistical analysis. The study explored Machine Learning models, including Recurrent Neural Networks and Long Short-Term Memory (LSTM) Networks. Hyperparameter optimization through GridSearch ensured optimal model performance. Training utilized two years of data, validated against the most recent six months. Forecast horizons ranged from 1 to 48 hours using a rolling window technique, revealing improved accuracy near the actual time horizon. Evaluation metrics such as Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) demonstrated predictive accuracy, with MAPEs showing a significant improvement. The results highlight the potential of data preprocessing and exogenous variable selection to enhance predictive accuracy, offering a pathway to optimize renewable energy integration in Spain's intraday market. These methods may also benefit other markets facing similar challenges.