Machine Learning-Based Dispatch Optimization of a Molten-Salt Concentrating Solar Power Plant Using Day-Ahead Electricity Price Forecasting
Description
This dataset contains the electricity-market, gas-market, power-system and economic time series used to develop and benchmark the end-to-end dispatch-optimization algorithm presented in the associated article, which couples a deep neural network for day-ahead electricity price forecasting with a genetic algorithm for revenue-maximizing thermal energy storage (TES) dispatch in a molten-salt parabolic-trough concentrating solar power plant operating in the Spanish day-ahead market. All series cover 1 January 2016 – 31 December 2021 and comprise: hourly day-ahead marginal prices published by OMIE (marginalpdbc series, one file per calendar day, provided both in the original non-standard format and as converted CSV); the daily natural-gas day price and traded volume from MIBGAS; the daily generation structure by technology from Red Eléctrica de España (ESIOS); daily Brent crude-oil prices; the IBEX 35 and IBEX 35 Energy indices; and meteorological variables retrieved from the AEMET OpenData API. These variables were screened as candidate predictors through a Pearson-correlation analysis against the day-ahead price, which retained the natural-gas day price and a categorical day-of-week indicator as the final predictor set. The dataset also includes the consolidated modelling file actually supplied to the forecasting model, the preprocessing and dataset-composition scripts used to convert, clean and merge the raw files, and the benchmark results of the genetic-algorithm dispatch module across the 32 test scenarios (season × gas-price regime × day type × irradiance pattern). The direct normal irradiance profiles and engineering parameters of the reference commercial plant are not included, as they were provided under a confidentiality agreement.
Files
Steps to reproduce
Raw OMIE files are downloaded from the public day-ahead price repository, one file per calendar day, in the native .1 format. DScreator.py converts each file to comma-separated values; DS_composer_ELE.py merges the daily files into a single time series. MIBGAS, ESIOS, Brent and IBEX series are downloaded as annual or full-period spreadsheets and reduced to the variables of interest (date, price, traded volume, generation by technology). AEMET data are retrieved through the OpenData API using the mining scripts provided. All series are conditioned to a common daily resolution and merged by date into the consolidated modelling dataset. Candidate predictors are screened by Pearson correlation against the day-ahead price; the retained predictors feed the TensorFlow forecasting model, trained on a three-way training/validation/test split. The dispatch module (PyGAD-based genetic algorithm) is executed per scenario with the plant model described in the article; scenario definitions and aggregated results are included in the benchmark folder.
Institutions
- Universidad Rey Juan CarlosMadrid, Madrid