This thesis is framed within the context of the European energy transition, which aims to achieve climate neutrality by 2050 and recognises the strategic role of photovoltaic (PV) and smart grid technologies in decarbonisation process. The research adopts a multidisciplinary approach to the topics of predictive maintenance (PdM) and intelligent energy management systems (EMS) applied to PV plants and hybrid AC/DC microgrids (MGs), integrating expertise in infrared thermography (IRT), artificial intelligence (AI) and physical-numerical modelling. The growing complexity of distributed systems and the variability of operating conditions make corrective or preventive maintenance approaches based on manual inspections and static thresholds inadequate. In this context, the combination of IRT and AI represents a high potential technological frontier, as it allows for rapid, non-destructive and automatable diagnostics, paving the way for predictive maintenance integrated with efficient EMS. The main goal is to build a standardised solution that can combine automatic fault diagnostics, state of health (SoH) assessment of components, and adaptive control of energy flows, helping to make next-gen power grids more reliable, efficient, and sustainable. The first part of the thesis analyses the physical principles of IRT and the relevant regulations, highlighting the measurement conditions and radiometric parameters necessary to ensure repeatability and traceability of results. An experimental thermographic campaign was conducted on the PV plant of the DICEAM Department of the Mediterranean University of Reggio Calabria, with the acquisition of IRT images. The images, annotated and balanced, enabled the creation of an original multi-class dataset, comprising six categories of defects: hotspots, bypass diode failure, substring faults, shading/soiling, glass damage and background. Starting from this dataset, a semantic segmentation model called Efficient Attentive U-Net was designed, based on U-Net encoder-decoder architectures with MobileNetV2 encoder, ASPP (Atrous Spatial Pyramid Pooling) modules for multi-scale aggregation, Attention Gates (AG) for spatial selection, and Squeeze-and-Excitation (SE) for adaptive channel weighting. This configuration made it possible to balance accuracy and computational lightness, achieving high performance with a reduced number of parameters and real-time inference, making the model appropriate for implementation on drones and edge devices. The proposed architecture allows for the accurate identification of segmented thermal maps which, in future developments, will enable the creation of predictive models capable of estimating the impact of faults on the electrical performance of the module. The methodology has been extended to the industrial domain, in particular to electrical distribution switchboards, through the development of a hybrid LSTM–U-Net (L-UN) model. The integration of temporal analysis of electrical signals and spatial segmentation of IRT images has allowed for the early detection of abnormal overheating and current overloads, anticipating fault conditions. The system, implemented on Raspberry Pi and Arduino platforms, showed excellent accuracy in predicting current absorption and automatically classifying thermal status, reducing unnecessary maintenance interventions. This multimodal application demonstrated the effectiveness of the PdM–AI–IRT paradigm even in complex industrial contexts, validating the possibility of creating intelligent distributed systems for the PdM of electrical components. The final part of the thesis broadens the field of investigation to the modelling of a hybrid AC/DC MG, including PV and wind turbine (WT) generators, a fuel cell system (PEMFC), battery storage (BESS) and heterogeneous AC/DC loads. The model, developed in MATLAB/Simulink©, was accompanied by a Mamdani-type fuzzy EMS, designed to coordinate production, storage, and consumption in real time. The EMS simultaneously integrates Power Quality (PQ) indicators (ΔV, Δf, THDV, TDD), state of charge (SOC) and dynamic tariff signals, pursuing dual technical and economic optimisation. Simulations have shown that the fuzzy EMS improves voltage stability, reduces harmonic distortion, increases self-consumption, and lowers daily operating costs. The main innovative contribution consists in the development of two complementary systems for intelligent PdM and EMS, each address specifics aspect of diagnostics and operational optimisation in hybrid MGs. The bidirectional information link between AI diagnostic modules and fuzzy controllers makes it possible, in the future, to adapt dispatch strategies according to the health status of generators and converters, evolving towards cognitive MGs capable of predicting, preventing, and optimising. In summary, the thesis provides a coherent and replicable framework that combines automatic thermographic analysis, physical modelling of energy systems and intelligent control, outlining a methodological path towards the digitalisation of maintenance and the predictive management of distributed networks. The results obtained confirm the validity of the proposed approach and lay the foundations for future research activities aimed at real-time AI-EMS integration, the adoption of digital twins for adaptive maintenance, and the use of interpretable neural models in accordance with IEC/IEEE standards for PQ and safety.

Questo lavoro di tesi si inserisce nel contesto della transizione energetica europea, che mira a raggiungere la neutralità climatica entro il 2050 e riconosce il ruolo strategico delle tecnologie fotovoltaiche (PV) e delle reti elettriche intelligenti nel processo di decarbonizzazione. La ricerca adotta un approccio multidisciplinare ai temi della manutenzione predittiva (PdM) e dei sistemi intelligenti di gestione dell'energia (EMS) applicati agli impianti PV e alle microgrid ibride AC/DC (MG), integrando competenze in termografia a infrarossi (IRT), intelligenza artificiale (AI) e modellazione fisico-numerica. La crescente complessità dei sistemi distribuiti e la variabilità delle condizioni operative rendono inadeguati gli approcci di manutenzione correttiva o preventiva basati su ispezioni manuali e soglie statiche. In questo contesto, la combinazione di IRT e AI rappresenta una frontiera tecnologica ad alto potenziale, in quanto consente una diagnostica rapida, non distruttiva e automatizzabile, aprendo la strada alla manutenzione predittiva integrata con un EMS efficiente. L'obiettivo principale è quello di costruire una soluzione standardizzata in grado di combinare la diagnostica automatica dei guasti, la valutazione dello stato di salute (SoH) dei componenti e il controllo adattivo dei flussi di energia, contribuendo a rendere le reti elettriche di nuova generazione più affidabili, efficienti e sostenibili. La prima parte della tesi analizza i principi fisici dell'IRT e le normative pertinenti, evidenziando le condizioni di misurazione e i parametri radiometrici necessari per garantire la ripetibilità e la tracciabilità dei risultati. È stata condotta una campagna termografica sperimentale sull'impianto PV del Dipartimento DICEAM dell'Università Mediterranea di Reggio Calabria, con l'acquisizione di immagini IRT. Le immagini, annotate e bilanciate, hanno consentito la creazione di un dataset multi-classe originale, comprendente sei categorie di difetti: hotspot, guasto dei diodi di bypass, guasti alle sottostringhe, ombreggiamento/sporcizia, danni al vetro e sfondo. Partendo da questo dataset, è stato progettato un modello di segmentazione semantica chiamato Efficient Attentive U-Net, basato su architetture U-Net encoder-decoder con encoder MobileNetV2, moduli ASPP (Atrous Spatial Pyramid Pooling) per l'aggregazione multiscala, Attention Gates (AG) per la selezione spaziale e Squeeze-and-Excitation (SE) per la ponderazione adattiva dei canali. Questa configurazione ha permesso di bilanciare accuratezza e leggerezza computazionale, ottenendo prestazioni elevate con un numero ridotto di parametri e inferenze in tempo reale, rendendo il modello adatto all'implementazione su droni e dispositivi edge. L'architettura proposta consente l'identificazione accurata delle mappe termiche segmentate che, in futuri sviluppi, consentiranno la creazione di modelli predittivi in grado di stimare l'impatto dei guasti sulle prestazioni elettriche del modulo. La metodologia è stata estesa al dominio industriale, in particolare ai quadri di distribuzione elettrica, attraverso lo sviluppo di un modello ibrido LSTM-U-Net (L-UN). L'integrazione dell'analisi temporale dei segnali elettrici e della segmentazione spaziale delle immagini IRT ha consentito il rilevamento precoce di surriscaldamenti anomali e sovraccarichi di corrente, anticipando le condizioni di guasto. Il sistema, implementato su piattaforme Raspberry Pi e Arduino, ha mostrato un'eccellente accuratezza nella previsione dell'assorbimento di corrente e nella classificazione automatica dello stato termico, riducendo gli interventi di manutenzione non necessari. Questa applicazione multimodale ha dimostrato l'efficacia del paradigma PdM-AI-IRT anche in contesti industriali complessi, convalidando la possibilità di creare sistemi distribuiti intelligenti per il PdM dei componenti elettrici. La parte finale della tesi amplia il campo di indagine alla modellizzazione di un MG ibrido AC/DC, che include generatori fotovoltaici e eolici (WT), un sistema a celle a combustibile (PEMFC), un sistema di accumulo a batteria (BESS) e carichi eterogenei AC/DC. Il modello, sviluppato in MATLAB/Simulink©, è stato accompagnato da un EMS fuzzy di tipo Mamdani, progettato per coordinare la produzione, lo stoccaggio e il consumo in tempo reale. L'EMS integra simultaneamente indicatori di qualità dell'energia (PQ) (ΔV, Δf, THDV, TDD), stato di carica (SOC) e segnali tariffari dinamici, perseguendo una doppia ottimizzazione tecnica ed economica. Le simulazioni hanno dimostrato che l'EMS fuzzy migliora la stabilità della tensione, riduce la distorsione armonica, aumenta l'autoconsumo e abbassa i costi operativi giornalieri. Il principale contributo innovativo consiste nello sviluppo di due sistemi complementari per il PdM e l'EMS intelligenti, ciascuno dei quali affronta aspetti specifici della diagnostica e dell'ottimizzazione operativa nei MG ibridi. Il collegamento bidirezionale di informazioni tra i moduli diagnostici AI e i controllori fuzzy rende possibile, in futuro, adattare le strategie di dispacciamento in base allo stato di salute dei generatori e dei convertitori, evolvendo verso MG cognitivi in grado di prevedere, prevenire e ottimizzare. In sintesi, la tesi fornisce un quadro coerente e replicabile che combina l'analisi termografica automatica, la modellizzazione fisica dei sistemi energetici e il controllo intelligente, delineando un percorso metodologico verso la digitalizzazione della manutenzione e la gestione predittiva delle reti distribuite. I risultati ottenuti confermano la validità dell'approccio proposto e gettano le basi per future attività di ricerca volte all'integrazione in tempo reale di AI-EMS, all'adozione di gemelli digitali per la manutenzione adattiva e all'uso di modelli neurali interpretabili in conformità con gli standard IEC/IEEE per la PQ e la sicurezza.

INFRARED THERMOGRAPHY AND ARTIFICIAL INTELLIGENCE FOR PREDICTIVE DIAGNOSTICS IN ELECTRICAL SYSTEMS From Photovoltaic Monitoring to Advanced Energy Management in Hybrid Microgrids / Pratticò, D.. - (2026 Apr 27).

INFRARED THERMOGRAPHY AND ARTIFICIAL INTELLIGENCE FOR PREDICTIVE DIAGNOSTICS IN ELECTRICAL SYSTEMS From Photovoltaic Monitoring to Advanced Energy Management in Hybrid Microgrids

Pratticò Danilo
2026-04-27

Abstract

This thesis is framed within the context of the European energy transition, which aims to achieve climate neutrality by 2050 and recognises the strategic role of photovoltaic (PV) and smart grid technologies in decarbonisation process. The research adopts a multidisciplinary approach to the topics of predictive maintenance (PdM) and intelligent energy management systems (EMS) applied to PV plants and hybrid AC/DC microgrids (MGs), integrating expertise in infrared thermography (IRT), artificial intelligence (AI) and physical-numerical modelling. The growing complexity of distributed systems and the variability of operating conditions make corrective or preventive maintenance approaches based on manual inspections and static thresholds inadequate. In this context, the combination of IRT and AI represents a high potential technological frontier, as it allows for rapid, non-destructive and automatable diagnostics, paving the way for predictive maintenance integrated with efficient EMS. The main goal is to build a standardised solution that can combine automatic fault diagnostics, state of health (SoH) assessment of components, and adaptive control of energy flows, helping to make next-gen power grids more reliable, efficient, and sustainable. The first part of the thesis analyses the physical principles of IRT and the relevant regulations, highlighting the measurement conditions and radiometric parameters necessary to ensure repeatability and traceability of results. An experimental thermographic campaign was conducted on the PV plant of the DICEAM Department of the Mediterranean University of Reggio Calabria, with the acquisition of IRT images. The images, annotated and balanced, enabled the creation of an original multi-class dataset, comprising six categories of defects: hotspots, bypass diode failure, substring faults, shading/soiling, glass damage and background. Starting from this dataset, a semantic segmentation model called Efficient Attentive U-Net was designed, based on U-Net encoder-decoder architectures with MobileNetV2 encoder, ASPP (Atrous Spatial Pyramid Pooling) modules for multi-scale aggregation, Attention Gates (AG) for spatial selection, and Squeeze-and-Excitation (SE) for adaptive channel weighting. This configuration made it possible to balance accuracy and computational lightness, achieving high performance with a reduced number of parameters and real-time inference, making the model appropriate for implementation on drones and edge devices. The proposed architecture allows for the accurate identification of segmented thermal maps which, in future developments, will enable the creation of predictive models capable of estimating the impact of faults on the electrical performance of the module. The methodology has been extended to the industrial domain, in particular to electrical distribution switchboards, through the development of a hybrid LSTM–U-Net (L-UN) model. The integration of temporal analysis of electrical signals and spatial segmentation of IRT images has allowed for the early detection of abnormal overheating and current overloads, anticipating fault conditions. The system, implemented on Raspberry Pi and Arduino platforms, showed excellent accuracy in predicting current absorption and automatically classifying thermal status, reducing unnecessary maintenance interventions. This multimodal application demonstrated the effectiveness of the PdM–AI–IRT paradigm even in complex industrial contexts, validating the possibility of creating intelligent distributed systems for the PdM of electrical components. The final part of the thesis broadens the field of investigation to the modelling of a hybrid AC/DC MG, including PV and wind turbine (WT) generators, a fuel cell system (PEMFC), battery storage (BESS) and heterogeneous AC/DC loads. The model, developed in MATLAB/Simulink©, was accompanied by a Mamdani-type fuzzy EMS, designed to coordinate production, storage, and consumption in real time. The EMS simultaneously integrates Power Quality (PQ) indicators (ΔV, Δf, THDV, TDD), state of charge (SOC) and dynamic tariff signals, pursuing dual technical and economic optimisation. Simulations have shown that the fuzzy EMS improves voltage stability, reduces harmonic distortion, increases self-consumption, and lowers daily operating costs. The main innovative contribution consists in the development of two complementary systems for intelligent PdM and EMS, each address specifics aspect of diagnostics and operational optimisation in hybrid MGs. The bidirectional information link between AI diagnostic modules and fuzzy controllers makes it possible, in the future, to adapt dispatch strategies according to the health status of generators and converters, evolving towards cognitive MGs capable of predicting, preventing, and optimising. In summary, the thesis provides a coherent and replicable framework that combines automatic thermographic analysis, physical modelling of energy systems and intelligent control, outlining a methodological path towards the digitalisation of maintenance and the predictive management of distributed networks. The results obtained confirm the validity of the proposed approach and lay the foundations for future research activities aimed at real-time AI-EMS integration, the adoption of digital twins for adaptive maintenance, and the use of interpretable neural models in accordance with IEC/IEEE standards for PQ and safety.
27-apr-2026
Settore ING-IND/31 - ELETTROTECNICA
Settore IIET-01/A - Elettrotecnica
LA FORESTA, Fabio
PIETRAFESA, Matilde Mariarosa Consolata
Doctoral Thesis
File in questo prodotto:
File Dimensione Formato  
PhD+Thesis_Pratticò+Danilo_XXXVIII+ciclo.pdf

accesso aperto

Tipologia: Tesi di dottorato
Licenza: Creative commons
Dimensione 3.41 MB
Formato Adobe PDF
3.41 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.12318/164906
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact