Portail HAL edf
Not a member yet
11029 research outputs found
Sort by
In-reactor stress relaxation of 3-point bending CW-316 samples: Experiments and modelling
No full textInternational audienceCW-</div
Apprentissage profond pour la caractérisation des séries temporelles de consommation électrique
No full textThe transition to low-carbon energy, reinforced by international agreements, demands greater flexibility in the electric grid to effectively integrate renewable energy sources.In this context, the widely deployed smart meters provide precise and time-stamped household electricity consumption data, serving as a key enabler for improving grid flexibility.However, their low temporal resolution (typically one reading every 30 minutes) limits their utility for fine-grained analysis, particularly in extracting detailed information on individual appliance consumption.This thesis introduces innovative deep learning-based approaches to analyze these time series and address the challenges of appliance detection and energy usage tracking in households.After establishing the theoretical foundations and reviewing the state of the art in time series analysis and Non-Intrusive Load Monitoring (NILM), this work first focuses on detecting appliance presence from low-frequency aggregated data.An evaluation of multiple classification techniques highlights the superiority of deep learning-based approaches.First, we investigate various time series classification methods and show that deep learning significantly outperforms traditional approaches for detecting whether a household owns a particular appliance. We then introduce ADF&TransApp, a detection framework leveraging Transformers pre-trained on large volumes of unlabeled data, thus making full use of the extensive consumption datasets available. Next, we address the localization of appliance activation periods.We propose CamAL, a weakly supervised approach combining convolutional networks and explainability techniques.A key innovation of our solution is its ability to be trained solely on appliance ownership information, significantly reducing the need for annotated data.To make this approach accessible, we developed DeviceScope, an interactive tool that highlights appliance signatures in consumption data, providing clearer insights for both consumers and electricity providers.Finally, to estimate individual appliance consumption, we introduce NILMFormer, a Transformer-based architecture incorporating a normalization mechanism tailored to the NILM problem.This design effectively handles the intrinsic variations in consumption data, addressing the challenge posed by their non-stationary nature.The work conducted in this thesis has led to the large-scale operational deployment of the proposed solutions, demonstrating their relevance for optimized energy management and their active contribution to the energy transition.La transition vers les énergies bas carbone, renforcée par les accords internationaux, impose une plus grande flexibilité du réseau électrique afin de gérer efficacement l’intégration des sources d’énergies renouvelables. Dans cette optique, les compteurs intelligents, désormais largement déployés, fournissent des données précises et horodatées sur la consommation électrique des foyers, constituant ainsi un levier essentiel pour améliorer cette flexibilité. Toutefois, leur faible résolution temporelle (typiquement un point relevé toutes les 30 minutes) limite l’exploitation fine de ces données, en particulier pour extraire des informations détaillées sur la consommation individuelle des appareils domestiques. Cette thèse propose des approches innovantes basées sur l’apprentissage profond afin d’analyser ces séries temporelles et de répondre aux enjeux de détection et de suivi des usages énergétiques des foyers.Après avoir établi les bases théoriques et présenté l’état de l’art des méthodes d’analyse de séries temporelles et de la surveillance non intrusive de la charge, ce travail se concentre d’abord sur des approches permettant de détecter la présence d’appareils à partir de données agrégées à faible fréquence. L’évaluation de plusieurs techniques de classification met en évidence la nette supériorité des approches basées sur l’apprentissage profond. Nous introduisons ainsi ADF&TransApp, un framework reposant sur des Transformers pré-entraînés, capable d’exploiter efficacement de larges volumes de données non annotées pour détecter avec précision les équipements présents dans les foyers.Dans un second temps, nous nous intéressons à la localisation des périodes d’activation des appareils. Pour cela, nous proposons CamAL, une approche faiblement supervisée combinant réseaux convolutifs et techniques d’explicabilité.L’originalité de cette méthode réside dans sa capacité à être entraînée uniquement à partir d’informations sur la possession des appareils, réduisant ainsi considérablement les besoins en annotations. Afin de rendre cette approche accessible, nous développons DeviceScope, un outil interactif qui met en évidence les signatures d’appareils dans les données de consommation, offrant ainsi une meilleure interprétation aux utilisateurs et aux fournisseurs d’électricité. Enfin, afin de restituer la consommation individuelle de chaque appareil, nous introduisons NILMFormer, une architecture basée sur les Transformers qui intègre un mécanisme spécifique de normalisation. Ce dispositif permet de gérer efficacement les variations intrinsèques des données de consommation et de relever ainsi le défi posé par leur nature non stationnaire.Les travaux réalisés dans cette thèse ont abouti à l’intégration opérationnelle à grande échelle des solutions proposées, démontrant ainsi leur pertinence pour une gestion optimisée de l’énergie et leur contribution active à la transition énergétique
Innovative Nb-Doped SnO 2 Electron Transport Layers Prepared by Atomic Layer Deposition for Enhanced Perovskite Solar Cells
No full textInternational audienceRapid advancements in perovskite solar cell (PSC) technology have highlighted the critical role of precise interface engineering in enhancing device stability and performance. Metal oxide-based electron transport layers (ETLs) prepared by atomic layer deposition (ALD) have emerged as promising candidates for improving PSC efficiency and stability. In this study, Nb-doped SnO2 (SnO2:Nb) thin films fabricated by ALD were employed as ETLs in n-i-p architecture PSCs. By leveraging the atomic-level control of ALD, the optoelectronic properties of SnO2:Nb thin films were finely tuned through controlled Nb doping, significantly influencing the photovoltaic (PV) performance of the devices in threshold behavior. PSCs incorporating SnO2:Nb thin film ETLs with low Nb atomic contents (≤0.2 at. %) exhibited improved performance, with an absolute increase in power conversion efficiency (PCE) of 0.93%, primarily due to increased Voc and Jsc values. Half-cell and extensive characterizations, including dark J–V curves, external and internal quantum efficiencies, impedance spectroscopy, and steady-state and time-resolved photoluminescence, were conducted to elucidate the effect of Nb doping. The perovskite layer was found to remain unaffected by ETL modification. The performance enhancement is attributed to the improved electrical properties of ETL thin films, leading to reduced series resistance and increased shunt resistance, as well as the reduction of interface defects, alteration of decay times, and favorable band alignments. These findings highlight the potential of ALD-processed SnO2-based ETLs for next-generation PSCs
Analysis and Optimization of a Liquid-vapor Thermohydraulic Model
No full textInternational audienceThis work presents a simplified model of compressible multiphase fluid flow in a heated porous medium. We first introduce the model and its numerical approximation using a linearized implicit finite volume scheme. We then propose a technique to accelerate the implicit scheme through a machine learning approach
Presymptomatic microRNA-based biomarker signatures for the prognosis of localized radiation injury in mice
No full textInternational audienceThe threat of nuclear or radiological events requires early diagnostic tools for radiation induced health effects. Localized radiation injuries (LRI) are severe outcomes of such events, characterized by a latent presymptomatic phase followed by symptom onset ranging from erythema and edema to ulceration and tissue necrosis. Early diagnosis is crucial for effective triage and adapted treatment, potentially through minimally invasive biomarkers including circulating microRNAs (miRNAs), which have been correlated with tissue injuries and radiation exposure, suggesting their potential in diagnosing LRI. In this study, we sought to identify early miRNA signatures for LRI severity prognosis before clinical symptoms appear. Using a mouse model of hindlimb irradiation at 0, 20, 40, or 80 Gy previously shown to lead to localized injuries of different severities, we performed broad-spectrum plasma miRNA profiling at two latency stages (day 1 and 7 post-irradiation). The identified candidate miRNAs were then challenged using two independent mouse cohorts to refine miRNA signatures. Through sparse partial least square discriminant analysis (sPLS-DA), signatures of 14 and 16 plasma miRNAs segregated animals according to dose groups at day 1 and day 7, respectively. Interestingly, these signatures shared 9 miRNAs, including miR-19a-3p, miR-93-5p, miR-140-3p, previously associated with inflammation, radiation response and tissue damage. In addition, the Bayesian latent variable modeling confirmed significant correlations between these prognostic miRNA signatures and day 14 clinical and functional outcomes from unrelated mice. This study identified plasma miRNA signatures that might be used throughout the latency phase for the prognosis of LRI severity. These results suggest miRNA profiling could be a powerful tool for early LRI diagnosis, thereby improving patient management and treatment outcomes in radiological emergency situations
Multi-output Gaussian processes for the reconstruction of homogenized cross-sections
No full textInternational audienceDeterministic nuclear reactor neutronics codes employing the prevalent two-step scheme often generate a substantial amount of intermediate data at the interface of their two subcodes, which can impede the overall performance of the software. The bulk of this data comprises “few-groups homogenized cross-sections” or HXS, which are stored as tabulated multivariate functions and interpolated inside the core code. A number of mathematical tools have been studied for this interpolation purpose over the years, but few meet all the challenging requirements of neutronics computation chains: extreme accuracy, low memory footprint, fast predictions. We here present a new framework to tackle this task, based on multi-output Gaussian processes (MOGP). These smooth and tunable bayesian regressors are able to model several quantities at once, and to capture the correlations between them – a key asset in the modeling of HXS’s, which we show to be highly similar from one another. Several models of this family are discussed, compared, adapted to the case of very numerous HXS’s, and their possible modeling choices are experimented on. These machine learning models enable us to interpolate HXS’s with improved accuracy compared to the current multilinear standard, using only a fraction of its training data – meaning that the amount of required precomputation is reduced by a factor of several dozens. They also necessitate an even smaller fraction of its storage requirements, preserve its reconstruction speed, and unlock new functionalities such as adaptive sampling and facilitated uncertainty quantification. We demonstrate the efficiency of this approach on a rich test case reproducing the VERA benchmark, proving in particular its scalability to datasets of millions of HXS’s
Multi-Agent Multi-Armed Bandits: application to EVs smart charging with grid constraints
No full textInternational audienceElectrification of energy uses with renewable sources is a major lever of decarbonization. To maximize the use of renewable energy, and tackle its variability and uncertainty, flexible entities such as electric vehicles can be used. This paper presents a scalable, decentralized multi-agent system, where eachelectric vehicle (EV) seeks the best instants to charge to satisfy its mobility needs, while favoring photovoltaic energy and avoiding congesting the electric network. Each EV makes autonomous de-cisions using information from its environment, using contextual multi-armed bandit algorithms based on Thompson sampling. Four types of bandit-based algorithms are compared to a baseline through simulations with 55 homogeneous EV agents, achieving a distance of only 14% from an optimal omniscient centralized algorithm. The system’s real-world practicality is demonstrated through minimal computational requirements (200 μs per EV per timestep) and limited information sharing, maintaining user privacy while effectively managing grid congestion
Machine Learning for Degradation Estimation of Hydro Generator Units
No full textInternational audienceIn the context of the ongoing electrification of energy usages and the integration of renewable energy capacity to address the challenges of climate change, several questions arise regarding the ability of the future electricity system to maintain the equilibrium between production and consumption, as new means for energy storage and transportation are developed. Historically in Europe, the task of balancing supply and demand for electricity has primarily fallen to thermal and hydro power plants, owing to their ability to swiftly adapt their power output on demand. However, as the average age of a hydropower plant in the continent is estimated to be 42 years, considering refurbishment dates, numerous challenges are foreseen concerning operational availability and security. Among various technologies considered, both proven and emerging, predictive maintenance holds particular significance for legacy powerplants
Génération d'une base de courriers électroniques synthétiques par des grands modèles de langue dans le domaine de la relation client
No full textNational audienceTraining AI systems for customer relations tasks is hindered by the high costs of manual data annotation and stringent privacy regulations. This article describes an approach using large language models to generate synthetic emails without incorporating real customer data. The aim is to train AI systems on these synthetic data, providing better privacy guarantees and minimizing the volume of manually annotated data. We describe the processing pipeline and its implementation, particularly the prompt creation phase, which captures the diversity of topics, styles, and types of personal data. Beyond generation, inserting fictitious entities into the text allows for the automatic annotation of an email similar to a real one. Evaluation results on a dataset of 1,600 emails indicate a promising approach for training AI systems while offering better guarantees in terms of privacy and regulatory compliance.Dans le domaine de la relation client, exploiter les données textuelles offre un levier essentiel pour développer des systèmes d'IA performants, mais présente d’importants défis de confidentialité et de conformité réglementaire, et nécessite souvent l'annotation manuelle et coûteuse de données. Cet article décrit une approche d'utilisation des grands modèles de langue pour générer des e‑mails synthétiques, sans intégrer de données client réelles. L'objectif est de permettre d'entrainer des systèmes d'IA sur ces données synthétiques, en offrant de meilleures garanties de protection de la vie privée, et en permettant de minimiser le volume de données manuellement annotées. Nous décrivons la chaîne de traitement et son implémentation notamment la phase de création de prompts, qui capture la diversité des sujets, des styles et les types de données personnelles. Au-delà de la génération, l’insertion d'entités fictives dans le texte permet de reformer un email automatiquement annoté similaire à un email réel. Les résultats d’évaluation sur un jeu de données de 1600 emails indiquent une piste prometteuse pour l'entraînement de systèmes d'IA tout en offrant de meilleures garanties du point de vue du respect de la vie privée et de la conformité réglementaire
What's inside your industrial black-box component? Let's analyze some micro-architectural signals!
No full textInternational audienceWith the growing complexity of systems, design phases increasingly rely on the interaction of several industrial actors. This makes it more difficult, especially at the hardware level, for an end-user to know what is being inserted at each stage, even for very specific needs, and can be a blocking point for revising the system later on.In terms of security, using only a high level of abstraction alone does not protect against several attacks or malicious acts that exploit the target’s low-level characteristics. Since only third parties know the implementation details of the component for which they are responsible, there is an increasing need for direct monitoring of signals coming from the micro-architecture to cover attacks targeting this layer. Suppliers may intentionally or not produce a component that is vulnerable to attacks against the micro-architecture. To detect attacks at this level, we propose a mechanism to extract a large set of signals and select to most relevant ones to study the behavior of industrial-type systems. These systems often have small processors with lightweight operating systems, sometimes with real-time constraints.To simulate various such systems, we have built an FPGA platform for continuous monitoring of the micro-architectural signals, based on LiteX, with different choices of parameters such as CPUs and peripherals. Our work extends the MATANA framework, which enables run-time detection of Cache Side-Channel and Return-Oriented Programming attacks. We are also extending the framework to support hardware trojans targeting industrial systems, with automated insertion tools. Experiments are designed for high bandwidth data transfer to a host computer