Soutenance de thèse de Luis Enrique GARCIA MARRERO

Jury Pannel :

M. Mickaël HILAIRET , de l 'Ecole Centrale de Nantes , Examinator

M. Nicolas PATIN, de l'Université Technologique de Compiègne, Previewer

M. João MARTINS, de l' Universidade Nova de Lisboa, Previewer

M. Eric MONMASSON , de CY Cergy Paris Université, Co-Supervisor

M. Giovanni PETRONE, de l' Università degli Studi di Salerno, Co-Supervisor

 

Keywords : Energy Management Systems (EMS), Non-Intrusive Load Monitoring (NILM), PV diagnostics 

Abstract :  

This thesis presents an integrated framework for demand- and supply-side intelligence in smart-building Energy Management Systems (EMS) designed for deployment on resource-constrained edge-computing platforms. The demand-side research focuses on appliance load monitoring using the cost-efficient Non-Intrusive Load Monitoring (NILM) approach. Current state-of-the-art NILM methods face significant limitations, including poor generalization to domain shifts and high computational requirements. This work begins with a theoretical study of convolutional neural network-based NILM architectures, quantifying performance degradation under domain shifts through a first-order Taylor expansion and identifying the primary error sources affecting generalization. To solve this practical limitation, a novel training-less NILM framework is developed, combining a probabilistic appliance state model, dynamic programming for sequential updates, a lightweight base-load estimation module, and a population-based incremental learning algorithm. The proposed method operates in real-time, is robust to domain shifts, and eliminates the need for abundant appliance-specific training data. The supply-side research addresses model-based PV diagnostics, with a focus on improving parameter identification under real-world operating conditions. The investigation begins with a multi-objective optimization framework for the joint identification of parameters in static and dynamic PV models, integrating current-voltage (I-V) curve data with electrochemical impedance spectroscopy (EIS) measurements to produce physically meaningful estimates. However, this approach is limited by the requirement for EIS hardware and the assumption of uniform operating conditions. To overcome these constraints, a self-adapting seven-parameter Double Single-Diode Model (D-SDM) is introduced, which estimates parameters using only I-V data. The method employs a robust error function to isolate valid curve segments and employs evolutionary algorithms for parameter fitting, ensuring stable and accurate estimation under real-world operating conditions, while reliably detecting degradation phenomena such as increases in series resistance. The NILM and PV diagnostic frameworks are integrated into a unified edge-computing platform capable of executing both tasks concurrently. All processing is performed locally, and results are delivered through a single user interface. Experimental validation confirms that the system operates with low latency and sustained performance, without exceeding hardware capacity. A commercialization framework is also proposed to support large-scale deployment, combining privacy-preserving local computation with premium cloud-based services, and targeting residential users, PV owners, and installers. This work is part of the SMARTGYsum (SMART Green energY Systems and bUsiness Models) research and training EU program.

 

Microsoft Teams

Rejoindre la réunion maintenant

Numéro de réunion : 327 350 659 885 72

Code secret : Jj3rK2UE