Soutenance de thèse Mohamad Moussa
 |
M. Mohamad Moussa soutiendra en anglais, en vue de l'obtention du grade de docteur ès sciences de la société, mention systèmes d'information, sa thèse intitulée: "Towards decentralised machine learning predictions for local energy communities" Date: Vendredi 28 novembre 2025 à 13h00 Lieu: sur Zoom |
Résumé :
The liberalization of the electricity market and the proliferation of new forms of electricity production and consumption are paving the way for new smart digital services. These services will involve a new generation of intelligent devices, called Grid Edge Devices (GEDs), designed to support electric systems and microgrids. Microgrids are localised energy systems capable of operating either connected to the main grid or independently in islanded mode.
Accurate prediction of energy consumption and production is essential for maintaining microgrid stability, and leveraging machine learning, whether through centralized models or privacy-preserving federated learning, enables adaptive, data-driven forecasting that supports efficient grid management.
In this thesis, we investigate and propose several federated learning (FL) approaches that leverage intelligence at the edge, ensuring data privacy. Unlike traditional machine learning models that require centralizing raw data for training, our methods enable edge devices to collaboratively train forecasting models locally, without transmitting raw data to a central server. This paradigm is particularly valuable in energy systems, where privacy, security, and communication constraints present significant operational challenges. Through extensive experiments using real-world energy consumption data provided by a Swiss energy company, we investigated different machine learning models for load forecasting.
Although XGBoost, DeepAR, and Transformer-based models were also in consideration, Long Short-Term Memory (LSTM)-based models were selected as the primary solution due to their favourable balance between predictive performance, and simplicity of adaptation to FL paradigms.
We explore and compare different FL architectures, including centralized - where learning models are collected and centralized into a global model, decentralized variants - where individuals models are aggregated only with a selected set of neighbouring node models, as well as continual learning extensions for adaptability in dynamic environments. Unsurprisingly, centralized FL achieves the highest predictive accuracy due to global model aggregation. However, from a decentralization perspective —where avoiding central dependence and excessive communication is desirable — our decentralized FL approaches demonstrate performance that closely approaches centralized FL, especially when communication is limited to a small number of similar neighbors. Furthermore, integrating continual learning through periodic model retraining, enables decentralized models to achieve performance that is comparable to, and in some cases even surpasses, multi-device or centralised configurations.
Overall, the results from this thesis show conclusively that decentralized federated learning mechanisms, especially using LSTM architecture and neighbor-aware collaboration, can serve as a viable and efficient alternative to centralised federated learning (FL) for energy forecasting in local energy communities.
Decentralized federated learning offers transformative potential for smart grids by enabling secure, privacy-preserving collaboration across distributed energy resources. By allowing edge devices to train models locally and share only model updates, decentralized FL eliminates the need to centralize sensitive energy data. This approach not only enhances data privacy and security but also reduces communication overhead and latency, making real-time forecasting and decision-making more efficient. In complex energy systems where data heterogeneity and operational constraints are prevalent, decentralized FL fosters scalable, adaptive intelligence that strengthens grid resilience, optimizes energy distribution, accelerates the adoption of renewable energy, and supports the transition toward more autonomous and sustainable energy infrastructures.
Jury de thèse :
- Prof. Nabil ABDENNADHER, co-directeur de thèse, HES-SO HEPIA, Suisse
- Prof. Raphael COUTURIER, co-directeur de thèse, Université Marie et Louis Pasteur, France
- Prof. Giovanna DI MARZO SERUGENDO, co-directrice de thèse, Université de Genève
- Prof. Jean-Henry MORIN, président du jury, Université de Genève
- Prof. Mokhtar BOZORG, HES-SO, HEIG-VD, Suisse
- Prof. Carlo FISCHIONE, KTH Royal Institute of Technology, Sweden