Titelangaben
Bremer, Onno ; Hanft, Alisa ; Zeilmann, Bernd ; Fischerauer, Gerhard:
Reliable AI-Driven PV Forecasting for Local, Resource-Constrained Systems.
2025
Veranstaltung: International Conference on Advanced Machine Learning and Data Science (AMLDS 2025)
, 19.–21. Juli 2025
, Tokyo, Japan.
(Veranstaltungsbeitrag: Kongress/Konferenz/Symposium/Tagung
,
Vortrag
)
Abstract
Global energy systems are undergoing a fundamental transformation, increasingly focusing on renewable energy sources and their efficient utilization. Model predictive control (MPC) in microgrids relies on accurate and reliable forecasts of photovoltaic (PV) power availability. Previous research on artificial intelligence (AI) for PV forecasting has often neglected real-world constraints. This study evaluates various forecasting approaches, including Convolutional Neural Networks (CNN), Support Vector Regression (SVR), and Nonlinear AutoRegressive models with eXogenous inputs (NARX), comparing them to PVlib, a physics-based analytical model. A key requirement for these models is efficient training and deployment on field-grade hardware, enabling localized, cloud-independent operation. Additionally, the important practical aspects of ease of commissioning and robustness under varying conditions were analyzed.
The evaluation used real-world data from rooftop PV systems (11–170 kWp) and a façade-mounted system, combined with weather forecasts from the German Weather Service (DWD). PVlib, while providing a strong analytical baseline, required extensive effort for deployment, especially in large installations. In contrast, AI models were optimized using Bayesian techniques to autonomously collect data and retrain periodically. Unlike PVlib, they were also able to account for module aging, soiling effects, and edge cases such as snow coverage on rooftop installations.
Feature sets included global, direct, and diffuse radiation, time of day, day of year, temperature, and wind speed. While irradiance data alone was sufficient for rooftop systems, incorporating time features (e.g., sinusoidal transformations) significantly improved accuracy for façade-mounted PV systems.
All AI models exhibited stable convergence, delivering superior forecasts compared to PVlib in both power curve accuracy and total daily energy yield, with significantly lower deployment effort. Among the evaluated models, SVR demonstrated the highest field applicability with excellent forecasting performance. These findings underscore AI’s critical role in enhancing the reliability and efficiency of PV power forecasting for future autonomous microgrids.
Weitere Angaben
Publikationsform: | Veranstaltungsbeitrag (Vortrag) |
---|---|
Begutachteter Beitrag: | Nein |
Keywords: | Photovoltaic power prediction; machine learning for energy systems; resource-constrained systems; real-world constraints in AI deployment; local and cloud-independent operation; Support Vector Regression (SVR); Convolutional Neural Networks (CNN); Nonlinear AutoRegressive models (NARX); Bayesian optimization;
Edge computing for PV forecasting; renewable energy integration; robust AI models for microgrids |
Institutionen der Universität: | Fakultäten Fakultäten > Fakultät für Ingenieurwissenschaften Fakultäten > Fakultät für Ingenieurwissenschaften > Lehrstuhl Mess- und Regeltechnik Fakultäten > Fakultät für Ingenieurwissenschaften > Lehrstuhl Mess- und Regeltechnik > Lehrstuhl Mess- und Regeltechnik - Univ.-Prof. Dr.-Ing. Gerhard Fischerauer Forschungseinrichtungen > Forschungsstellen > Zentrum für Energietechnik - ZET |
Titel an der UBT entstanden: | Ja |
Themengebiete aus DDC: | 600 Technik, Medizin, angewandte Wissenschaften > 620 Ingenieurwissenschaften |
Eingestellt am: | 01 Aug 2025 07:05 |
Letzte Änderung: | 01 Aug 2025 07:12 |
URI: | https://eref.uni-bayreuth.de/id/eprint/94159 |