Announcement
Access to the virtual meeting platform is now closed.
Access to the virtual meeting platform is now closed.
Presentation Details
| Short-term Data-driven Forecasting of Post-curtailment Aggregate Photovoltaic Generation in Cyprus Viken Davidian1, 2, Panagiotis Herodotou1, 3, Demetris Marangis1, 3, Rogiros Tapakis2, Stavros Afxentis2, Constantinos Alexandrou2, Stelios Chiras2, George Makrides1, 2, George E.Georghiou1, 2. 1PHAETHON Centre of Excellence (CoE) for Intelligent, Efficient and Sustainable Energy Solutions, Uni, Nicosia, Cyprus.2PV Technology Lab, Department of Electrical and Computer Engineering, University of Cyprus, Nicosia, Cyprus.3Department of Mechanical and Manufacturing Engineering, University of Cyprus, Nicosia, Cyprus.4Transmission System Operator of Cyprus, Nicosia, Cyprus |
Abstract
Due to the isolated nature and infrastructure limitations of the Cyprus power system, a substantial amount of renewable energy sources (RES) is curtailed. Advancements in the field of artificial intelligence (AI) have led to the development of data-driven forecasting methodologies that can be utilized for photovoltaic (PV) generation forecasting and help reduce operational uncertainty and mitigate curtailment. To address this issue, this paper presents a multi-output Extreme Gradient Boosting (XGBoost) based model capable of providing day-ahead (D+1) to four-day-ahead (D+4) forecasts of post-curtailment aggregate photovoltaic (PV) generation taking into account spatial variability of meteorological conditions, seasonal load patterns, and lagged operational fuel cost of conventional units. The proposed model was trained, and its hyperparameters were optimized twice: once with historical weather observations (observation-trained) and once with historical numerical weather predictions, or NWPs (NWP-trained). Three scenarios were evaluated against a persistence baseline: (i) the observation-trained model predicting with ideal weather forecasts, (ii) the observation-trained model predicting with historical NWPs, and (iii) the NWP-trained model predicting with historical NWPs. In operational conditions, the NWP-trained model outperformed both the observation-trained model and the persistence baseline across all horizons (D+1 to D+4), achieving skill scores of 0.436, 0.425, 0.381, and 0.366 relative to the baseline. Finally, it achieved a mean daily Root Mean Square Error (RMSE) of 38.261 MW, demonstrating robustness across a range of weather conditions and indicating readiness for operational deployment.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.
No part of this publication may be reproduced, distributed, or transmitted in any form or by any means, including photocopying, recording, or other electronic or mechanical methods, without the prior written permission of the author.