Summary of the research project: The Upper Rhine region has been undergoing a profound transformation of its local energy system for several years. More than 850,000 households or small collective structures now produce their own solar electricity while remaining connected to the distribution network, forming what are known as prosumers. This increasingly growing phenomenon generates new and poorly anticipated constraints on distribution networks: intermittent injections, consumption peaks that are difficult to predict, multiplication of batteries of very different natures. Faced with this reality, the management tools available on the market are too often compartmentalized by brand or country, and none of them really takes advantage of what all these installations collectively represent. 

This PhD offer is part of the European FAIR'nRG Interreg Upper Rhine project, coordinated by CESI in partnership  with higher education institutions and energy stakeholders from France, Germany, and Switzerland. FAIR'nRG aims to develop a multifunctional platform for prosumers in the Upper Rhine, based on distributed artificial intelligence approaches based on a sovereign regional cloud infrastructure. Within this ambitious and collaborative scientific framework, the doctoral research work will be structured around three major axes. 

The first axis aims to design and optimize a robust and efficient federated learning architecture, capable of managing the heterogeneity of data and equipment (inverters, batteries, sensors) present on the trinational territory. Our work has led us to develop methods for selecting and grouping customers to accelerate the convergence of AI models in highly heterogeneous contexts both in terms of data and computational performance [1, 2]. We are now interested in the application of clustering techniques to efficiently aggregate information from prosumers with diverse profiles. The objective is then to maintain a certain degree of specialization of customer models, while capitalizing on a global common information. The whole will have to be functional on constrained devices on the client side and will be able to absorb the increase in infrastructure load. 

 The second axis will focus on the development of dynamic optimization strategies. It will involve creating algorithms capable of integrating continuous flow variables, such as local weather forecasts, consumption profiles, and network price signals, to maximize self-consumption and the lifespan of storage systems [3]. These will be parameterized by the aggregated information and characterized by the AI models mentioned above. These approaches, based on heuristics or metaheuristics, will be designed to respond to the diversity of prosumer profiles according to a progressive spectrum: from a manual intervention by the user to a complete automation of the energy system. This deliberately graduated structure of the problem will allow the PhD student to build a solid and well-bounded contribution, articulating personalization of responses and scaling. Finally, the third axis will explore incremental learning, allowing models to refine themselves continuously from incoming new data flows without the need for massive and centralized retraining and make the scaling of the entire infrastructure more reliable. 

In the final phase of the thesis, the performance modules developed will be confronted with the reality of the field. The PhD student will participate in the deployment of the FAIR'nRG pilot, involving several dozen real prosumers across France, Germany and Switzerland. This experimental validation will evaluate the ability of the models to adapt to various operational scenarios and demonstrate the added value of collective and distributed intelligence. 

The objectives of the thesis are listed below:  

• Design of efficient methods for customer aggregation in a heterogeneous federated AI context (clustering). 
• Design of efficient models to maintain customer specialization in a federated AI context. 
• Development of multi-profile optimization algorithms based on hybrid heuristics or meta-heuristics powered by federated AI. 
• Integration of AI models into FAIR'nRG's multi-scale infrastructure and scientific valorization. 
• Writing of the thesis manuscript, presentation of the results, defense. 

Expected results: The PhD student will contribute to the production of an original software architecture integrating distributed energy optimization models, deployed and validated in real conditions within the framework of the FAIR'nRG program. The scientific deliverables will include publications in peer-reviewed journals, papers in international conferences, as well as a functional prototype integrated into the project platform. The thesis will directly participate in the construction of a trinational observatory of distributed storage capacities in the Upper Rhine. 

Funding category: Financement de l'Union européenne



PHD Country: France