University of Western Macedonia
Investigation of optimal strategies for photovoltaic system operation
Abstract
dc:descriptionAs the installed capacity of Photovoltaic (PV) systems increases, strategies for their optimal operation are essential to be employed. The optimal operation of PV systems depends on their normal operation and profitability and requires not only efficient Operation and Maintenance (O&M) strategies but also their competitive participation in electricity markets. Traditional O&M approaches—time-based Preventive Maintenance (PvM), condition-based PvM, and predictive methods—primarily focus on detecting faults or degradation without addressing the increasingly demanding question of how corrective and replacement actions should be prioritized. As the installed PV capacity grows, so does the number of issues requiring attention, intensifying the operational burden on O&M contractors. This complexity underscores the need for Prescriptive Maintenance (RxM) methodologies, which are designed to prioritize actions and allocate human resources optimally. This thesis aims to investigate methods for the optimal operation of PV system focusing on: i) the development of a comprehensive RxM methodology tailored to PV systems O&M and ii) the development of a holistic forecasting framework to support PV systems’ participation in electricity market, considering the European Target Model rules. The RxM methodology integrates three key components: a holistic point-forecasting model, a multi-layered PvM model with AI-based Digital Twin (DT) integration. More specifically, the forecasting model introduces a Long Short-Term Memory (LSTM) model trained using a hybrid Adam–Unified Particle Swarm Optimization (UPSO) algorithm. The hybrid optimizer leverages Adam’s fast convergence and UPSO’s global search capability, enabling more accurate power predictions under limited numerical weather prediction inputs. Across multiple datasets, input scenarios, and benchmark optimizers, the hybrid approach outperforms conventional methods in 66.67% of cases, achieving normalized mean absolute error reductions of up to 16.17% compared to persistence and up to 6.46% relative to Adam alone. Additionally, the development of PvM model was based on the integration of AI-based DTs to enhance its fault detection (FD) and fault classification (FC) accuracy. Neural Network and LSTM-based DTs are trained to simulate PV performance using only meteorological inputs, achieving R² values above 0.99. These simulations are combined with measured data and fed into six classifiers. Results demonstrate that DT-enhanced FD and FC significantly improve classification performance, with string-level modeling yielding superior accuracy and certain classifiers, such as Support Vector Machines, NN, and Extreme Learning Machines, benefiting the most from DT inputs. Building upon the forecasting and PvM layers, the RxM model was formulated as a Mixed-Integer Nonlinear Programming (MINLP) that minimizes total O&M costs, including labor, fuel, and energy-loss costs. Thirteen scenarios and multiple system topologies are used to benchmark the RxM model against alternative scheduling schemes. The proposed approach yields a consistent 4.5–17.06% reduction in total O&M costs, outperforming rule-based and single-objective strategies. In parallel, a holistic forecasting framework is developed to support PV systems’ participation in both the Day-Ahead Market (DAM) and the Intra-Day Market (IDM) in line with European Target Model rules. Four forecasting models—covering DAM, and the three auctions of IDM—were developed using Convolutional Neural Network (CNN) and Transformer architectures. Results across six datasets show that Transformer-based models reduce Mean Absolute Error by 27.94–32.45% compared to CNNs, with the forecasting model of the second and the third auction of IDM providing the most substantial improvements in forecast accuracy. Overall, the thesis delivers two complementary frameworks: an RxM methodology that enables cost-optimal O&M planning, and a DAM/IDM-compliant forecasting model that enhances market participation. Together, these contributions advance the operational reliability, economic performance, and market integration of large-scale PV systems.
Degree
thesis:*- Grantor dc:publisher
- University of Western Macedonia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Kothona, Despoina
- Κοθώνα, Δέσποινα
Subjects
dc:subject × 24- Φωτοβολταϊκά συστήματα
- Λειτουργία και συντήρηση φωτοβολταϊκών συστημάτων
- Τεχνητή νοημοσύνη
- Πρόβλεψη φωτοβολταϊκής παραγωγής
- Ψηφιακά δίδυμα
- Μοντέλα βελτιστοποίησης
- Αγορά ηλεκτρικής ενέργειας
- Photovoltaic systems
- Operation and maintenance of photovoltaic systems
- Artificial intelligence
- Photovoltaic power forecasting
- Digital twins
- Optimization models
- Electricity market
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Μαθηματικά
- Έλεγχος και Βελτιστοποίηση
- Μοντελοποίηση και Προσομοίωση
- Computer and Information Sciences
- Natural Sciences
- Mathematics
- Control and Optimization
- Modeling and Simulation
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/10442/hedi/61174