Exploring machine learning algorithms for managing a system of renewable energy and storage assets
This thesis explores the integration of machine learning forecasting techniques with an optimization model designed to minimize the Levelized Cost of Hydrogen (LCOH) in an off-grid renewable energy system. The study evaluates the impact of forecast accuracy on resource allocation, battery utilizatio...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/428625 |
| Acceso en línea: | https://hdl.handle.net/2117/428625 |
| Access Level: | acceso abierto |
| Palabra clave: | Machine learning -- Forecasting Hydrogen as fuel Renewable energy sources Aprenentatge automàtic -- Previsió Hidrogen com a combustible Energies renovables Àrees temàtiques de la UPC::Energies |
| Sumario: | This thesis explores the integration of machine learning forecasting techniques with an optimization model designed to minimize the Levelized Cost of Hydrogen (LCOH) in an off-grid renewable energy system. The study evaluates the impact of forecast accuracy on resource allocation, battery utilization, and hydrogen production, highlighting the importance of reliable forecasting to ensure optimal planning and operational decisions. The system under study consists of a solar PV farm, a wind farm, a Li-ion battery, and an alkaline electrolyzer. The optimization model incorporates a detailed representation of the electrolyzer, based on Ulleberg’s mathematical model. Three supervised learning models are employed to forecast renewable electricity production: Linear Regression (LR), Support Vector Regression (SVR) and Long Short-Term Memory (LSTM). Together with a seasonal Na¨ ıve model developed as a benchmark to facilitate the comparison when analyzing the results. Results show that while all machine learning models outperform the Na¨ ıve benchmark, they present smoother predictions that lead to worse results in the optimization problem. The LR model achieves the lowest mean absolute error (MAE) and root mean square error (RMSE) but tends to overestimate electricity production, reducing the need for energy storage. The LSTM model developed performs better than the rest of models when capturing high and low generation peaks, but has higher computational costs. The SVR model, in contrast, shows the highest forecasting errors and signs of over-fitting. Overall, the project highlights the importance of accurate prediction in renewable energy systems, and the significance of not only relying on accuracy metrics such as the RMSE or MAE to assess the forecasting performance of machine learning algorithms. Finally, offering a foundation for future research to contribute to the large-scale production of hydrogen. |
|---|