Thèse des Réseaux Électriques aux Aéronefs Plus Électriques une Approche Unifiée pour la Détection Précoce des Instabilités Oscillatoires dans les Systèmes Énergétiques Dominés par les Convertisse H/F - Doctorat.Gouv.Fr
- Toulouse - 31
- CDD
- Doctorat.Gouv.Fr
Les missions du poste
Établissement : Institut National Polytechnique de Toulouse École doctorale : GEETS - Génie Electrique Electronique,Télécommunications et Santé : du système au nanosystème Laboratoire de recherche : LGP - Laboratoire Génie de Production Direction de la thèse : Mohamed KOUKI ORCID 0000000283476260 Début de la thèse : 2027-09-01 Date limite de candidature : 2026-11-23T23:59:59 Les systèmes énergétiques électriques modernes connaissent une transformation profonde sous l'effet de la pénétration croissante des modules de puissances (convertissuers,...), des sources d'énergie renouvelable à base d'onduleurs, des commandes numériques à haute bande passante, des réseaux de distribution en courant continu (DC) et de l'électrification des charges. Cette évolution améliore la flexibilité, l'efficacité et le potentiel de décarbonation des systèmes électriques, mais modifie également leur comportement dynamique et favorise l'apparition de nouvelles formes d'instabilités oscillatoires. Celles-ci peuvent se manifester sous la forme d'oscillations interzones à basse fréquence, d'oscillations forcées, d'oscillations subsynchrones, d'interactions entre convertisseurs et leurs systèmes de commande, d'oscillations des bus DC ou encore d'interactions entre les sources et les charges.
Les perturbations récentes observées sur les réseaux électriques ont montré que ces phénomènes oscillatoires peuvent précéder des événements en cascade susceptibles de conduire à des défaillances majeures, notamment en présence d'un faible amortissement, de limitations des dispositifs de contrôle ou de conditions d'exploitation dominées par les convertisseurs. Dans ce contexte, le développement d'outils d'alerte précoce, capables de détecter en temps réel les précurseurs de ces instabilités tout en restant interprétables et généralisables, constitue un enjeu scientifique et industriel majeur.
Cette projet de thèse propose le développement d'un cadre unifié d'Intelligence Artificielle Informée par la Physique (Physics-Informed Artificial Intelligence) pour la détection précoce, la classification, la localisation et l'évaluation du risque des instabilités oscillatoires dans les systèmes énergétiques dominés par les convertisseurs. L'application principale portera sur la surveillance des réseaux de transport d'électricité à partir des unités de mesure de phase (PMU) et des systèmes WAMS, afin d'identifier les oscillations de très basses fréquences et les oscillations subsynchrones avant qu'elles ne provoquent des perturbations critiques.
La méthodologie sera ensuite étendue aux micro-réseaux et aux systèmes électriques d'aéronefs plus électriques (More-Electric Aircraft), afin d'évaluer sa capacité à détecter les interactions locales entre convertisseurs, les oscillations des bus DC, les instabilités liées aux charges à puissance constante et les phénomènes critiques des réseaux embarqués.
L'approche combinera des techniques de traitement du signal, d'analyse temps-fréquence, d'apprentissage automatique, de modèles d'apprentissage profond temporels guidés par des indicateurs de stabilité. Contrairement aux approches d'intelligence artificielle de type 'boîte noire', le modèle intégrera explicitement des informations physiques telles que la fréquence des oscillations, l'évolution de l'amortissement, l'énergie modale et spectrale, les oppositions de phase entre zones, les marges d'impédance, les limites des convertisseurs et les incertitudes de mesure.
Le résultat attendu est le développement d'un « Indice de Risque d'Instabilité Oscillatoire » permettant de caractériser et de classifier en temps réel l'état du système selon plusieurs niveaux de criticité, allant d'un fonctionnement normal à des situations présentant un risque élevé d'instabilité. Cette recherche contribuera au développement d'un outil d'alerte précoce interprétable, robuste et transférable, favorisant la prévention des blackouts dans les réseaux électriques, ainsi que le renforcement de la résilience des micro-réseaux et la sécurisation de l'électrification des aéronefs de nouvelle génération. Electrical power systems are undergoing a profound transformation from rotational-machine-dominated architectures to converter-dominated architectures driven by the large-scale integration of renewable energy sources, power electronic converters, and electrified loads. In conventional power systems, synchronous generators inherently provide essential stability services, including rotational inertia, electromechanical damping, short-circuit strength, and frequency support. By contrast, modern converter-dominated systems rely primarily on fast digital control algorithms, power electronic interfaces, and communication-based coordination. Consequently, their dynamic behavior is governed by control loops, phase-locked loops, current-limiting mechanisms, DC-link dynamics, grid-forming and grid-following converter interactions, as well as strong source-load coupling. These characteristics introduce new stability challenges and significantly modify the mechanisms responsible for oscillatory phenomena.
This paradigm shift is evident across three major application domains investigated in this PhD thesis.
The first application domain concerns large-scale transmission power systems. Modern interconnected transmission networks are experiencing an unprecedented penetration of converter-interfaced renewable energy resources, including wind farms, photovoltaic power plants, battery energy storage systems, and HVDC transmission lines. Although these technologies improve system flexibility and sustainability, they also reduce the effective system inertia and modify the damping characteristics traditionally provided by synchronous generators. Consequently, the power system becomes more susceptible to a wide range of oscillatory phenomena, including low-frequency electromechanical oscillations, sub-synchronous oscillations, and converter-driven control interactions. Since these oscillations may represent precursors to cascading failures or widespread blackouts, their early detection using synchronized measurements provided by PMUs/WAMS is essential for enhancing real-time situational awareness and improving system resilience.
The second application domain focuses on converter-dominated microgrids. These systems typically integrate distributed renewable generation, battery energy storage, grid-forming and grid-following inverters, DC/DC converters, and constant power loads operating in both AC and DC networks. Owing to their reduced inertia and high penetration of power electronic interfaces, microgrids exhibit dynamic behaviors that differ substantially from those of conventional power systems. Their stability is strongly influenced by inverter interactions, droop-control parameters, DC-bus voltage dynamics, converter bandwidths, and the negative incremental impedance characteristics associated with constant power loads. These complex interactions may lead to poorly damped oscillations or even instability, particularly under islanded operating conditions.
The third application domain addresses More-Electric Aircraft (MEA). The aviation industry is progressively replacing conventional hydraulic and pneumatic subsystems with electrically driven actuators, high-power converters, high-voltage DC distribution networks, and advanced embedded power management systems. From a systems perspective, these electrical architectures can be regarded as highly dynamic, safety-critical isolated microgrids. Their reliable operation requires continuous monitoring of converter interactions, source-load coupling, variable-frequency generation, and DC-bus voltage dynamics. Detecting oscillatory instabilities at an early stage is therefore essential to ensure system reliability, operational safety, and fault-tolerant performance.
Despite significant differences in scale, operating environment, and application objectives, these three domains share a common scientific challenge: the reliable early detection and characterization of oscillatory instability signatures in converter-dominated electrical systems using measurement-based monitoring techniques. Developing unified data-driven methods capable of identifying the onset of oscillatory instabilities across these diverse applications constitutes the central research objective of this PhD thesis.
The scientific challenge of this thesis lies in the development of a unified, reliable, and interpretable AI-based methodology capable of detecting early oscillatory instability precursors across three different converter-dominated systems: transmission power grids, microgrids, and More-Electric Aircraft electrical networks. Although these systems differ in scale, voltage level, measurement infrastructure, and operational constraints, they share common dynamic challenges related to low damping, converter interactions, source-load coupling, and multi-timescale oscillatory behavior.
Challenge 1: Early Detection of Weak and Multi-Timescale Oscillatory Phenomena - The first challenge is to detect oscillatory instability before it becomes severe or visible through conventional alarms. In transmission grids, this concerns very-low-frequency and sub-synchronous oscillations that may appear before large disturbances or cascading events. In microgrids, the challenge is to detect local oscillations caused by inverter interactions, islanding events, droop-control conflicts, or DC-bus dynamics. In More-Electric Aircraft systems, the difficulty is to detect fast DC-bus oscillations, source-load interactions, and electromechanical coupling under highly dynamic operating conditions. These oscillations occur over different frequency ranges and time scales. Therefore, a single fixed-threshold or single-frequency detection method is not sufficient. The thesis must develop adaptive signal-processing and AI-based methods capable of extracting weak oscillatory signatures from heterogeneous measurements.
Challenge 2: Physics-Informed AI Under Limited, Noisy, and Heterogeneous Data - The second challenge is the lack of large, labelled, and homogeneous instability datasets. In transmission grids, real blackout or near-blackout events are rare and PMU data may be incomplete, noisy, or confidential. In microgrids, experimental data are limited and strongly dependent on the tested topology, converter parameters, and load configuration. In aircraft electrical systems, real instability data are even more difficult to access because of safety certification and industrial confidentiality constraints. Pure non-explainable AI (black-box AI) is therefore not sufficient. The AI framework must be guided by physical indicators such as oscillation frequency, damping trend, modal energy, spectral energy growth, inter-area phase opposition, impedance stability margin, DC-bus stiffness, converter current limits, and source-load interaction indicators. This physics-informed approach is necessary to improve robustness, reduce false alarms, and make the AI decisions interpretable for operators and engineers.
Challenge 3: Transferability Across Systems with Different Scales and Constraints- The third challenge is cross-domain transferability. A transmission grid, a laboratory microgrid, and a More-Electric Aircraft electrical system have very different architectures, voltage levels, sampling rates, and safety requirements. A model trained directly on one system cannot be applied without adaptation. However, these systems share common instability mechanisms associated with converter-dominated dynamics (like damping degradation, oscillation energy growth, impedance interaction, control-loop coupling, and source-load instability). The scientific challenge is therefore to identify transferable features and learning representations that can be adapted from the main PMU-based grid application to microgrids and then to aircraft electrical systems. The thesis will address this challenge through transfer learning, domain adaptation, physics-informed feature extraction, and explainable AI. The objective is not to use exactly the same trained model for all applications, but to develop a unified methodology that can be adapted to different converter-dominated systems while preserving physical interpretability and early-warning capability.
State of the Art and Research Gap
Oscillatory instability has long been a central topic in power-system stability studies, traditionally associated with synchronous-machine dynamics, local modes, inter-area modes, and forced oscillations, usually analyzed through eigenvalue analysis, Prony analysis, matrix pencil methods, damping-ratio estimation, and modal observability [1, 2]. With the increasing penetration of inverter-based resources, HVDC links, renewable power plants, storage systems, and fast converter controls, the nature of oscillatory phenomena is changing: oscillations may now originate from grid-following synchronization, grid-forming interactions, weak-grid conditions, converter-control loops, DC-link dynamics, and reduced natural damping [7-10], [28, 29]. PMU/WAMS-based monitoring has therefore become essential for observing wide-area dynamics, since PMUs provide synchronized measurements of voltage and current phasors, frequency, ROCOF, active power, and reactive power according to synchrophasor standards [3, 4]. Recent European grid events, including the 2025 Iberian disturbance, further highlight the need for early detection of very-low-frequency and sub-synchronous oscillations before they contribute to voltage instability, cascading disconnections, or blackout conditions [5, 6]. However, PMU-based detection remains challenging when oscillations are weak, low-frequency, masked by noise, or mixed with nonstationary operating changes. To address these limitations, recent research increasingly combines signal decomposition, time-frequency analysis, and explainable AI. Methods such as Empirical Mode Decomposition (EMD), Ensemble EMD, CEEMDAN, Variational Mode Decomposition, wavelet analysis, Hilbert transform, and short-time Fourier transform can extract hidden oscillatory modes, instantaneous frequency, amplitude envelope, modal energy, and damping-related features from nonstationary signals [11-13]. Machine learning classifiers such as Decision Trees, k-Nearest Neighbors, Support Vector Machines, Random Forests, and Artificial Neural Networks can then be used for event identification, while deep learning models such as CNNs, LSTM/GRU networks, Transformers, autoencoders, and graph neural networks can capture temporal, spectral, and topology-aware patterns [14-20]. A preliminary submitted study on PMU-based sustained sub-synchronous oscillation detection has already shown the relevance of combining EMD with supervised classifiers, where SVM and ANN achieved strong performance and generalization on unseen European datasets [27]. Nevertheless, many AI-based approaches still lack physical interpretability, statistical validation, and robustness under limited labelled data. Beyond transmission grids, similar oscillatory stability problems appear in converter-dominated microgrids and More-Electric Aircraft electrical systems. In microgrids, instability may result from inverter interactions, droop-control conflicts, islanding events, DC-bus dynamics, and constant-power-load behavior, where negative incremental impedance can reduce damping and lead to oscillations [21,22]. In More-Electric Aircraft, electrical architectures include generators, rectifiers, DC/DC converters, active front-end converters, 28 V DC buses, 270 V or ±270/540 V DC buses, high-frequency or variable-frequency AC networks, actuators, avionics, and dynamic loads [23-26]. The literature shows that aircraft electrical systems may suffer from phase/frequency synchronization issues, impedance-related instabilities, electromechanical interactions between generators and drivetrains, and source-load oscillations [24-26]. These systems are different in scale and constraints, but they share common converter-dominated instability mechanisms (like damping degradation, oscillation-energy growth, impedance interactions, control-loop coupling, and source-load instability). The main research gap is therefore the absence of a unified physics-informed and transferable AI framework able to detect early oscillatory instability precursors across transmission grids, microgrids, and aircraft electrical systems while providing interpretable alarms and a practical risk index.
Research Hypothesis
The central hypothesis of this thesis is that the onset of oscillatory instabilities in converter-dominated electrical systems is preceded by measurable temporal, spectral, spatial, and physical signatures that can be reliably detected using Physics-Informed AI models applied to synchronized and local electrical measurements.
This hypothesis is divided into four sub-hypotheses:
a) Very-low-frequency and sub-synchronous oscillatory modes can be extracted from low-rate PMU/WAMS data using adaptive signal decomposition and time-frequency analysis.
b) AI classifiers can distinguish critical oscillatory events from normal transients when trained on physically meaningful features rather than only raw measurements.
c) Physical indicators such as modal energy, damping trend, spectral growth, phase opposition, impedance margin, source-load stability margin, and converter operating limits improve AI robustness and interpretability.
d) A methodology developed for PMU-based transmission-grid monitoring can be adapted to microgrids and More-Electric Aircraft electrical systems using transfer learning, domain adaptation, and application-specific features.
The main objective is to develop a Physics-Informed AI framework for early detection of oscillatory instabilities in converter-dominated energy systems, with a main validation on PMU-based transmission grids and transferability assessment on microgrids and More-Electric Aircraft systems.
The specific objectives are:
O1) Develop an AI-driven and physics-informed framework for oscillatory instability monitoring by establishing a unified taxonomy of oscillation phenomena, extracting early oscillatory signatures from PMU/WAMS data, developing AI models for detection, classification, localization, and severity assessment, integrating physics-informed stability indicators, and constructing an Oscillatory Instability Risk Index for early-warning applications.
O2) Validate and evaluate the proposed framework by comparing classical machine learning, deep learning, and unsupervised approaches, and by validating the methodology using both simulated and real transmission-grid data.
O3) Demonstrate the generalization, interpretability, and applicability of the framework by transferring the methodology to converter-dominated microgrids and More-Electric Aircraft electrical systems, while providing interpretable decision-support tools for transmission system operators, microgrid engineers, and aircraft electrical-system designers.
The proposed methodology is organized into three main research stages from the definition of the scientific problem to the development and validation of a unified Physics-Informed Artificial Intelligence framework for the early detection of oscillatory instabilities. Although presented sequentially, the literature review and methodological improvements will continue throughout the thesis as new challenges and results emerge.
Stage 1 - State of the Art and Problem Definition
The first stage aims to establish the scientific foundations of the research. It includes a comprehensive review of oscillatory instability phenomena in modern electrical systems, covering transmission power grids, converter-dominated microgrids, and More-Electric Aircraft electrical systems. Particular attention will be paid to low-frequency, sub-synchronous, converter-induced, DC-bus, and source-load oscillations, together with existing monitoring techniques based on PMU/WAMS measurements, signal-processing methods, artificial intelligence, and physics-informed approaches. This stage will identify the current research gaps, define the datasets, operating scenarios, and instability classes, and formulate the research questions and objectives that will guide the remainder of the thesis.
Stage 2 - Development of the Physics-Informed AI Framework
The second stage constitutes the core contribution of the thesis. Measurement data will first be collected, synchronized, cleaned, and preprocessed to build consistent datasets for AI-based analysis. Advanced signal decomposition and time-frequency analysis techniques, including EMD and its variants, will then be employed to extract oscillatory modes and reveal hidden dynamic behaviors. Based on these signals, a set of physics-informed temporal, spectral, modal, and converter-related features will be constructed to preserve the physical interpretation of the monitored phenomena. These features will be integrated into artificial intelligence models combining classical machine learning, deep learning, and, where appropriate, graph-based learning techniques for oscillation detection, classification, localization, and severity assessment. Finally, explainable AI methods will be incorporated to improve model transparency and support the development of an Oscillatory Instability Risk Index for early-warning applications.
Stage 3 - Validation, Generalization, and Performance Assessment
The final stage focuses on validating and assessing the proposed framework. The transmission power grid will serve as the main validation case using both simulated and real PMU/WAMS datasets. The generalization capability of the methodology will then be evaluated on converter-dominated microgrids and More-Electric Aircraft electrical systems to investigate its adaptability to different converter-rich environments. The proposed framework will be assessed using conventional classification metrics (accuracy, precision, recall, F1-score, ROC-AUC, PR-AUC), statistical validation methods, computational performance, robustness, and explainability. Comparative analyses against conventional signal-processing and machine-learning approaches will demonstrate the effectiveness, reliability, and transferability of the proposed Physics-Informed AI framework across multiple electrical systems.
Les perturbations récentes observées sur les réseaux électriques ont montré que ces phénomènes oscillatoires peuvent précéder des événements en cascade susceptibles de conduire à des défaillances majeures, notamment en présence d'un faible amortissement, de limitations des dispositifs de contrôle ou de conditions d'exploitation dominées par les convertisseurs. Dans ce contexte, le développement d'outils d'alerte précoce, capables de détecter en temps réel les précurseurs de ces instabilités tout en restant interprétables et généralisables, constitue un enjeu scientifique et industriel majeur.
Cette projet de thèse propose le développement d'un cadre unifié d'Intelligence Artificielle Informée par la Physique (Physics-Informed Artificial Intelligence) pour la détection précoce, la classification, la localisation et l'évaluation du risque des instabilités oscillatoires dans les systèmes énergétiques dominés par les convertisseurs. L'application principale portera sur la surveillance des réseaux de transport d'électricité à partir des unités de mesure de phase (PMU) et des systèmes WAMS, afin d'identifier les oscillations de très basses fréquences et les oscillations subsynchrones avant qu'elles ne provoquent des perturbations critiques.
La méthodologie sera ensuite étendue aux micro-réseaux et aux systèmes électriques d'aéronefs plus électriques (More-Electric Aircraft), afin d'évaluer sa capacité à détecter les interactions locales entre convertisseurs, les oscillations des bus DC, les instabilités liées aux charges à puissance constante et les phénomènes critiques des réseaux embarqués.
L'approche combinera des techniques de traitement du signal, d'analyse temps-fréquence, d'apprentissage automatique, de modèles d'apprentissage profond temporels guidés par des indicateurs de stabilité. Contrairement aux approches d'intelligence artificielle de type 'boîte noire', le modèle intégrera explicitement des informations physiques telles que la fréquence des oscillations, l'évolution de l'amortissement, l'énergie modale et spectrale, les oppositions de phase entre zones, les marges d'impédance, les limites des convertisseurs et les incertitudes de mesure.
Le résultat attendu est le développement d'un « Indice de Risque d'Instabilité Oscillatoire » permettant de caractériser et de classifier en temps réel l'état du système selon plusieurs niveaux de criticité, allant d'un fonctionnement normal à des situations présentant un risque élevé d'instabilité. Cette recherche contribuera au développement d'un outil d'alerte précoce interprétable, robuste et transférable, favorisant la prévention des blackouts dans les réseaux électriques, ainsi que le renforcement de la résilience des micro-réseaux et la sécurisation de l'électrification des aéronefs de nouvelle génération. Electrical power systems are undergoing a profound transformation from rotational-machine-dominated architectures to converter-dominated architectures driven by the large-scale integration of renewable energy sources, power electronic converters, and electrified loads. In conventional power systems, synchronous generators inherently provide essential stability services, including rotational inertia, electromechanical damping, short-circuit strength, and frequency support. By contrast, modern converter-dominated systems rely primarily on fast digital control algorithms, power electronic interfaces, and communication-based coordination. Consequently, their dynamic behavior is governed by control loops, phase-locked loops, current-limiting mechanisms, DC-link dynamics, grid-forming and grid-following converter interactions, as well as strong source-load coupling. These characteristics introduce new stability challenges and significantly modify the mechanisms responsible for oscillatory phenomena.
This paradigm shift is evident across three major application domains investigated in this PhD thesis.
The first application domain concerns large-scale transmission power systems. Modern interconnected transmission networks are experiencing an unprecedented penetration of converter-interfaced renewable energy resources, including wind farms, photovoltaic power plants, battery energy storage systems, and HVDC transmission lines. Although these technologies improve system flexibility and sustainability, they also reduce the effective system inertia and modify the damping characteristics traditionally provided by synchronous generators. Consequently, the power system becomes more susceptible to a wide range of oscillatory phenomena, including low-frequency electromechanical oscillations, sub-synchronous oscillations, and converter-driven control interactions. Since these oscillations may represent precursors to cascading failures or widespread blackouts, their early detection using synchronized measurements provided by PMUs/WAMS is essential for enhancing real-time situational awareness and improving system resilience.
The second application domain focuses on converter-dominated microgrids. These systems typically integrate distributed renewable generation, battery energy storage, grid-forming and grid-following inverters, DC/DC converters, and constant power loads operating in both AC and DC networks. Owing to their reduced inertia and high penetration of power electronic interfaces, microgrids exhibit dynamic behaviors that differ substantially from those of conventional power systems. Their stability is strongly influenced by inverter interactions, droop-control parameters, DC-bus voltage dynamics, converter bandwidths, and the negative incremental impedance characteristics associated with constant power loads. These complex interactions may lead to poorly damped oscillations or even instability, particularly under islanded operating conditions.
The third application domain addresses More-Electric Aircraft (MEA). The aviation industry is progressively replacing conventional hydraulic and pneumatic subsystems with electrically driven actuators, high-power converters, high-voltage DC distribution networks, and advanced embedded power management systems. From a systems perspective, these electrical architectures can be regarded as highly dynamic, safety-critical isolated microgrids. Their reliable operation requires continuous monitoring of converter interactions, source-load coupling, variable-frequency generation, and DC-bus voltage dynamics. Detecting oscillatory instabilities at an early stage is therefore essential to ensure system reliability, operational safety, and fault-tolerant performance.
Despite significant differences in scale, operating environment, and application objectives, these three domains share a common scientific challenge: the reliable early detection and characterization of oscillatory instability signatures in converter-dominated electrical systems using measurement-based monitoring techniques. Developing unified data-driven methods capable of identifying the onset of oscillatory instabilities across these diverse applications constitutes the central research objective of this PhD thesis.
The scientific challenge of this thesis lies in the development of a unified, reliable, and interpretable AI-based methodology capable of detecting early oscillatory instability precursors across three different converter-dominated systems: transmission power grids, microgrids, and More-Electric Aircraft electrical networks. Although these systems differ in scale, voltage level, measurement infrastructure, and operational constraints, they share common dynamic challenges related to low damping, converter interactions, source-load coupling, and multi-timescale oscillatory behavior.
Challenge 1: Early Detection of Weak and Multi-Timescale Oscillatory Phenomena - The first challenge is to detect oscillatory instability before it becomes severe or visible through conventional alarms. In transmission grids, this concerns very-low-frequency and sub-synchronous oscillations that may appear before large disturbances or cascading events. In microgrids, the challenge is to detect local oscillations caused by inverter interactions, islanding events, droop-control conflicts, or DC-bus dynamics. In More-Electric Aircraft systems, the difficulty is to detect fast DC-bus oscillations, source-load interactions, and electromechanical coupling under highly dynamic operating conditions. These oscillations occur over different frequency ranges and time scales. Therefore, a single fixed-threshold or single-frequency detection method is not sufficient. The thesis must develop adaptive signal-processing and AI-based methods capable of extracting weak oscillatory signatures from heterogeneous measurements.
Challenge 2: Physics-Informed AI Under Limited, Noisy, and Heterogeneous Data - The second challenge is the lack of large, labelled, and homogeneous instability datasets. In transmission grids, real blackout or near-blackout events are rare and PMU data may be incomplete, noisy, or confidential. In microgrids, experimental data are limited and strongly dependent on the tested topology, converter parameters, and load configuration. In aircraft electrical systems, real instability data are even more difficult to access because of safety certification and industrial confidentiality constraints. Pure non-explainable AI (black-box AI) is therefore not sufficient. The AI framework must be guided by physical indicators such as oscillation frequency, damping trend, modal energy, spectral energy growth, inter-area phase opposition, impedance stability margin, DC-bus stiffness, converter current limits, and source-load interaction indicators. This physics-informed approach is necessary to improve robustness, reduce false alarms, and make the AI decisions interpretable for operators and engineers.
Challenge 3: Transferability Across Systems with Different Scales and Constraints- The third challenge is cross-domain transferability. A transmission grid, a laboratory microgrid, and a More-Electric Aircraft electrical system have very different architectures, voltage levels, sampling rates, and safety requirements. A model trained directly on one system cannot be applied without adaptation. However, these systems share common instability mechanisms associated with converter-dominated dynamics (like damping degradation, oscillation energy growth, impedance interaction, control-loop coupling, and source-load instability). The scientific challenge is therefore to identify transferable features and learning representations that can be adapted from the main PMU-based grid application to microgrids and then to aircraft electrical systems. The thesis will address this challenge through transfer learning, domain adaptation, physics-informed feature extraction, and explainable AI. The objective is not to use exactly the same trained model for all applications, but to develop a unified methodology that can be adapted to different converter-dominated systems while preserving physical interpretability and early-warning capability.
State of the Art and Research Gap
Oscillatory instability has long been a central topic in power-system stability studies, traditionally associated with synchronous-machine dynamics, local modes, inter-area modes, and forced oscillations, usually analyzed through eigenvalue analysis, Prony analysis, matrix pencil methods, damping-ratio estimation, and modal observability [1, 2]. With the increasing penetration of inverter-based resources, HVDC links, renewable power plants, storage systems, and fast converter controls, the nature of oscillatory phenomena is changing: oscillations may now originate from grid-following synchronization, grid-forming interactions, weak-grid conditions, converter-control loops, DC-link dynamics, and reduced natural damping [7-10], [28, 29]. PMU/WAMS-based monitoring has therefore become essential for observing wide-area dynamics, since PMUs provide synchronized measurements of voltage and current phasors, frequency, ROCOF, active power, and reactive power according to synchrophasor standards [3, 4]. Recent European grid events, including the 2025 Iberian disturbance, further highlight the need for early detection of very-low-frequency and sub-synchronous oscillations before they contribute to voltage instability, cascading disconnections, or blackout conditions [5, 6]. However, PMU-based detection remains challenging when oscillations are weak, low-frequency, masked by noise, or mixed with nonstationary operating changes. To address these limitations, recent research increasingly combines signal decomposition, time-frequency analysis, and explainable AI. Methods such as Empirical Mode Decomposition (EMD), Ensemble EMD, CEEMDAN, Variational Mode Decomposition, wavelet analysis, Hilbert transform, and short-time Fourier transform can extract hidden oscillatory modes, instantaneous frequency, amplitude envelope, modal energy, and damping-related features from nonstationary signals [11-13]. Machine learning classifiers such as Decision Trees, k-Nearest Neighbors, Support Vector Machines, Random Forests, and Artificial Neural Networks can then be used for event identification, while deep learning models such as CNNs, LSTM/GRU networks, Transformers, autoencoders, and graph neural networks can capture temporal, spectral, and topology-aware patterns [14-20]. A preliminary submitted study on PMU-based sustained sub-synchronous oscillation detection has already shown the relevance of combining EMD with supervised classifiers, where SVM and ANN achieved strong performance and generalization on unseen European datasets [27]. Nevertheless, many AI-based approaches still lack physical interpretability, statistical validation, and robustness under limited labelled data. Beyond transmission grids, similar oscillatory stability problems appear in converter-dominated microgrids and More-Electric Aircraft electrical systems. In microgrids, instability may result from inverter interactions, droop-control conflicts, islanding events, DC-bus dynamics, and constant-power-load behavior, where negative incremental impedance can reduce damping and lead to oscillations [21,22]. In More-Electric Aircraft, electrical architectures include generators, rectifiers, DC/DC converters, active front-end converters, 28 V DC buses, 270 V or ±270/540 V DC buses, high-frequency or variable-frequency AC networks, actuators, avionics, and dynamic loads [23-26]. The literature shows that aircraft electrical systems may suffer from phase/frequency synchronization issues, impedance-related instabilities, electromechanical interactions between generators and drivetrains, and source-load oscillations [24-26]. These systems are different in scale and constraints, but they share common converter-dominated instability mechanisms (like damping degradation, oscillation-energy growth, impedance interactions, control-loop coupling, and source-load instability). The main research gap is therefore the absence of a unified physics-informed and transferable AI framework able to detect early oscillatory instability precursors across transmission grids, microgrids, and aircraft electrical systems while providing interpretable alarms and a practical risk index.
Research Hypothesis
The central hypothesis of this thesis is that the onset of oscillatory instabilities in converter-dominated electrical systems is preceded by measurable temporal, spectral, spatial, and physical signatures that can be reliably detected using Physics-Informed AI models applied to synchronized and local electrical measurements.
This hypothesis is divided into four sub-hypotheses:
a) Very-low-frequency and sub-synchronous oscillatory modes can be extracted from low-rate PMU/WAMS data using adaptive signal decomposition and time-frequency analysis.
b) AI classifiers can distinguish critical oscillatory events from normal transients when trained on physically meaningful features rather than only raw measurements.
c) Physical indicators such as modal energy, damping trend, spectral growth, phase opposition, impedance margin, source-load stability margin, and converter operating limits improve AI robustness and interpretability.
d) A methodology developed for PMU-based transmission-grid monitoring can be adapted to microgrids and More-Electric Aircraft electrical systems using transfer learning, domain adaptation, and application-specific features.
The main objective is to develop a Physics-Informed AI framework for early detection of oscillatory instabilities in converter-dominated energy systems, with a main validation on PMU-based transmission grids and transferability assessment on microgrids and More-Electric Aircraft systems.
The specific objectives are:
O1) Develop an AI-driven and physics-informed framework for oscillatory instability monitoring by establishing a unified taxonomy of oscillation phenomena, extracting early oscillatory signatures from PMU/WAMS data, developing AI models for detection, classification, localization, and severity assessment, integrating physics-informed stability indicators, and constructing an Oscillatory Instability Risk Index for early-warning applications.
O2) Validate and evaluate the proposed framework by comparing classical machine learning, deep learning, and unsupervised approaches, and by validating the methodology using both simulated and real transmission-grid data.
O3) Demonstrate the generalization, interpretability, and applicability of the framework by transferring the methodology to converter-dominated microgrids and More-Electric Aircraft electrical systems, while providing interpretable decision-support tools for transmission system operators, microgrid engineers, and aircraft electrical-system designers.
The proposed methodology is organized into three main research stages from the definition of the scientific problem to the development and validation of a unified Physics-Informed Artificial Intelligence framework for the early detection of oscillatory instabilities. Although presented sequentially, the literature review and methodological improvements will continue throughout the thesis as new challenges and results emerge.
Stage 1 - State of the Art and Problem Definition
The first stage aims to establish the scientific foundations of the research. It includes a comprehensive review of oscillatory instability phenomena in modern electrical systems, covering transmission power grids, converter-dominated microgrids, and More-Electric Aircraft electrical systems. Particular attention will be paid to low-frequency, sub-synchronous, converter-induced, DC-bus, and source-load oscillations, together with existing monitoring techniques based on PMU/WAMS measurements, signal-processing methods, artificial intelligence, and physics-informed approaches. This stage will identify the current research gaps, define the datasets, operating scenarios, and instability classes, and formulate the research questions and objectives that will guide the remainder of the thesis.
Stage 2 - Development of the Physics-Informed AI Framework
The second stage constitutes the core contribution of the thesis. Measurement data will first be collected, synchronized, cleaned, and preprocessed to build consistent datasets for AI-based analysis. Advanced signal decomposition and time-frequency analysis techniques, including EMD and its variants, will then be employed to extract oscillatory modes and reveal hidden dynamic behaviors. Based on these signals, a set of physics-informed temporal, spectral, modal, and converter-related features will be constructed to preserve the physical interpretation of the monitored phenomena. These features will be integrated into artificial intelligence models combining classical machine learning, deep learning, and, where appropriate, graph-based learning techniques for oscillation detection, classification, localization, and severity assessment. Finally, explainable AI methods will be incorporated to improve model transparency and support the development of an Oscillatory Instability Risk Index for early-warning applications.
Stage 3 - Validation, Generalization, and Performance Assessment
The final stage focuses on validating and assessing the proposed framework. The transmission power grid will serve as the main validation case using both simulated and real PMU/WAMS datasets. The generalization capability of the methodology will then be evaluated on converter-dominated microgrids and More-Electric Aircraft electrical systems to investigate its adaptability to different converter-rich environments. The proposed framework will be assessed using conventional classification metrics (accuracy, precision, recall, F1-score, ROC-AUC, PR-AUC), statistical validation methods, computational performance, robustness, and explainability. Comparative analyses against conventional signal-processing and machine-learning approaches will demonstrate the effectiveness, reliability, and transferability of the proposed Physics-Informed AI framework across multiple electrical systems.
Le profil recherché
Le candidat doit être titulaire d'un Master ou diplôme équivalent en génie électrique, systèmes électriques, électronique de puissance, automatique, systèmes énergétiques, ou dans un domaine proche.
Un candidat idéal doit avoir des bases solides en systèmes électriques, dynamique et stabilité des réseaux électriques, électronique de puissance, systèmes à base de convertisseurs et automatique est souhaitée. Des connaissances en intégration des énergies renouvelables, micro-réseaux, réseaux à courant continu ou systèmes électriques des aéronefs plus électriques.
Le candidat doit aussi disposer de bonnes compétences en modélisation numérique, simulation, traitement du signal et analyse de données, avec une maîtrise de MATLAB/Simulink et/ou Python. Une première expérience en apprentissage automatique, apprentissage profond, intelligence artificielle ou intelligence artificielle informée par la physique sera appréciée, sans être indispensable si le candidat possède de solides bases en génie électrique et une forte motivation pour développer ces compétences au cours de la thèse.
Des connaissances en analyse temps-fréquence, identification de systèmes, optimisation, analyse de données PMU/WAMS ou évaluation des oscillations et de la stabilité seront également appréciées.
Le candidat doit faire preuve de curiosité scientifique, autonomie, capacités d'analyse et de résolution de problèmes, ainsi que d'une aptitude à travailler dans un environnement de recherche multidisciplinaire et international combinant génie électrique, intelligence artificielle et méthodes numériques avancées.
Dans le cadre de la collaboration scientifique entre UTTOP en France et l'Université de Sheffield au Royaume-Uni, le candidat devra être en mesure de travailler efficacement dans un environnement de recherche impliquant les équipes des deux établissements. Il devra également être disponible et disposé à effectuer une ou plusieurs périodes de mobilité de recherche à l'Université de Sheffield au cours de la thèse, en fonction des besoins scientifiques et des activités prévues dans le projet.
Un bon niveau d'anglais, à l'écrit comme à l'oral, est requis.
Application link: https://edd-projets.utoulouse.fr/
Un candidat idéal doit avoir des bases solides en systèmes électriques, dynamique et stabilité des réseaux électriques, électronique de puissance, systèmes à base de convertisseurs et automatique est souhaitée. Des connaissances en intégration des énergies renouvelables, micro-réseaux, réseaux à courant continu ou systèmes électriques des aéronefs plus électriques.
Le candidat doit aussi disposer de bonnes compétences en modélisation numérique, simulation, traitement du signal et analyse de données, avec une maîtrise de MATLAB/Simulink et/ou Python. Une première expérience en apprentissage automatique, apprentissage profond, intelligence artificielle ou intelligence artificielle informée par la physique sera appréciée, sans être indispensable si le candidat possède de solides bases en génie électrique et une forte motivation pour développer ces compétences au cours de la thèse.
Des connaissances en analyse temps-fréquence, identification de systèmes, optimisation, analyse de données PMU/WAMS ou évaluation des oscillations et de la stabilité seront également appréciées.
Le candidat doit faire preuve de curiosité scientifique, autonomie, capacités d'analyse et de résolution de problèmes, ainsi que d'une aptitude à travailler dans un environnement de recherche multidisciplinaire et international combinant génie électrique, intelligence artificielle et méthodes numériques avancées.
Dans le cadre de la collaboration scientifique entre UTTOP en France et l'Université de Sheffield au Royaume-Uni, le candidat devra être en mesure de travailler efficacement dans un environnement de recherche impliquant les équipes des deux établissements. Il devra également être disponible et disposé à effectuer une ou plusieurs périodes de mobilité de recherche à l'Université de Sheffield au cours de la thèse, en fonction des besoins scientifiques et des activités prévues dans le projet.
Un bon niveau d'anglais, à l'écrit comme à l'oral, est requis.
Application link: https://edd-projets.utoulouse.fr/