The production of electricity through photovoltaic (PV) systems represents one of the most significant drivers of the energy transition towards renewable sources. In the European Union alone, it accounts for 9.1% of the total gross energy production, and this share is expected to grow in the coming years given the Union’s climate policies. However, these systems are subject to numerous faults and failures in their electrical components, which can cause both damage and losses in generated power, thus leading to economic costs. To identify these faults, automated data-driven tools based on Machine Learning (ML) models are often implemented, thanks to their high responsiveness and scalability on large data sets, while maintaining generally low overheads. This work aims to investigate whether the use of physics can improve the automatic approach to Fault Detection and Diagnosis (FDD) by exploring existing limitations regarding the prediction and detection accuracy of current models when applied to plants different from those used during training. Physics is leveraged both in defining a dimensionless approach to the problem—thus making the method agnostic to the knowledge of the panels’ electrical parameters—and through the implementation of novel Physics-informed Neural Network models. In these models, the physical equations governing the operation of photovoltaic cells are incorporated into the training loss function, thereby improving reliability at inference time on panels not seen during training. These data are provided by Ricerca sul Sistema Energetico (RSE), which artificially induced and collected real faults. The research conducted demonstrates how the use of these hybrid data-physical approaches provides generalization benefits when dealing with test data from PV plants that differ form those used to train the model compared to a classical data-driven only pipeline.
La produzione di energia elettrica mediante impianti fotovoltaici (PV) rappresenta uno dei driver più significativi della transizione energetica verso fonti energetiche rinnovabili. Nella sola Unione Europea, essa rappresenta il 9.1% della produzione totale lorda di energia, ed è destinata a crescere nei prossimi anni viste le politiche climatiche dell’Unione. Tuttavia, questi sistemi sono soggetti a numerosi guasti e e defezioni alla componentistica elettrica, tale da causare sia danneggiamenti che perdite di potenza prodotta, comportando così costi economici. Per identificare questi guasti, vengono spesso implementati strumenti automatici basati sui dati attraverso modelli di Machine Learning (ML), grazie alla loro elevata prontezza e scalabilità su grandi moli di dati, pur mantenendo costi generalmente contenuti. Questo lavoro cerca di comprendere se l’utilizzo della fisica possa migliorare l’approccio automatico alla Fault Detection and Diagnosis (FDD) indagando i limiti esistenti in merito alla bontà di prediction e detection di modelli esistenti su tettoie diverse da quelle di addestramento. La fisica viene chiamata in causa sia nella definizione di un approccio dimensionless al problema, che quindi rende l’approccio agnostico rispetto alla conoscenza dei parametri elettrici dei pannelli, sia mediante l’implementazione di nuovi modelli di tipo Physics-informed Neural Network, dove l’equazioni fisiche modellizzanti il funzionamento della cella fotovoltaica sono incorporate nella loss function di training, migliorando di conseguenza l’affidabilità ad inference time su pannelli mai visti in fase di addestramento. Questi dati sono realizzati dall’azienda Ricerca sul Sistema Energetico (RSE), che ha indotto artificialmente e collezionato guasti reali. La ricerca condotta dimostra come l’uso di questi approcci basati sull’ibridazione fisica-dati offra una maggior generalizzazione ad inference time su dati provenienti da impianti fotovoltaici diversi da quelli utilizzati per addestrare il modello rispetto a una classica pipeline basata esclusivamente sui dati.
A Physical Approach to improve Generalization in automated Fault Detection and Diagnosis pipelines for Photovoltaic Plants
PANAROTTO, MATTEO
2024/2025
Abstract
The production of electricity through photovoltaic (PV) systems represents one of the most significant drivers of the energy transition towards renewable sources. In the European Union alone, it accounts for 9.1% of the total gross energy production, and this share is expected to grow in the coming years given the Union’s climate policies. However, these systems are subject to numerous faults and failures in their electrical components, which can cause both damage and losses in generated power, thus leading to economic costs. To identify these faults, automated data-driven tools based on Machine Learning (ML) models are often implemented, thanks to their high responsiveness and scalability on large data sets, while maintaining generally low overheads. This work aims to investigate whether the use of physics can improve the automatic approach to Fault Detection and Diagnosis (FDD) by exploring existing limitations regarding the prediction and detection accuracy of current models when applied to plants different from those used during training. Physics is leveraged both in defining a dimensionless approach to the problem—thus making the method agnostic to the knowledge of the panels’ electrical parameters—and through the implementation of novel Physics-informed Neural Network models. In these models, the physical equations governing the operation of photovoltaic cells are incorporated into the training loss function, thereby improving reliability at inference time on panels not seen during training. These data are provided by Ricerca sul Sistema Energetico (RSE), which artificially induced and collected real faults. The research conducted demonstrates how the use of these hybrid data-physical approaches provides generalization benefits when dealing with test data from PV plants that differ form those used to train the model compared to a classical data-driven only pipeline.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Panarotto_Thesis.pdf
solo utenti autorizzati a partire dal 19/11/2026
Descrizione: Elaborato di tesi
Dimensione
48.13 MB
Formato
Adobe PDF
|
48.13 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Panarotto_ExecutiveSummary.pdf
solo utenti autorizzati a partire dal 19/11/2026
Descrizione: Executive Summary della tesi
Dimensione
508.46 kB
Formato
Adobe PDF
|
508.46 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/246865