Prof. Gilles Notton

University of Corsica Pasquale Paoli, France

Speech Title

Better Energy Management through Better Sharing: The Concept of Energy Community

Biography

Gilles Notton is full professor in renewable energy systems with more than 37 years of experience. He received his PhD Degree and his "Accreditation to Supervise Researches" degree in Energy Engineering from the University of Corsica Pasquale Paoli in 1992 and 2002, respectively. The main objective of his research is to make intermittent and random renewable energies more easily controllable and usable by energy managers.

Gilles Notton's main research interests include renewable energy potential estimation and forecasting, hybrid renewable systems, energy management in smart electrical grid and particular energy situation of island territories.

Gilles Notton took part in multiple international research projects in collaboration with industrial partners and public parties such as the Horizon 2020 Tilos Project and COST Action TU1205.

In addition, he has published around 150 articles in international scientific journals and has more than 200 presentations in international conferences. He has participated in the development of several scientific books. He created and was responsible for an international scientific network on renewable energy between France and Eastern and Central European Countries.

He is associate-editor for the journal Renewable Energy, Elsevier.

Academic Profile: Google Scholar  |  ORCID: 0000-0002-6267-9632

Abstract

Faced with the continuous increase in electricity consumption, it is necessary to find sustainable, efficient and inexpensive energy-supply solutions. This necessarily involves the development of renewable energies which, due to their intermittency and unpredictability, are not easily integrated into the electricity grid. Emerging solutions, made possible by the development of new information technologies and by changes in regulations, have emerged. Among them, collective renewable self-consumption, coupled with smart micro-grids and the development of storage, is a serious and rapidly developing candidate. The objective of this talk is to present the principle of this collective self-consumption, its advantages as well as its constraints and the wide range of research that this sector opens up in the fields of energy, electrical engineering, new technologies without forgetting the sociological and legal aspects.

Prof. Zhengwei Li

Tongji University, China

Speech Title

Study of Transfer Reinforcement Learning in Air Conditioning Cooling Water System

Biography

Li Zhengwei, born in October 1981, holds a Ph.D. and serves as an Associate Professor and Master's Supervisor at the School of Mechanical and Energy Engineering, Tongji University. He received his Bachelor's degree in Building Environment and Equipment Engineering from Xi'an Jiaotong University in 2003, followed by a Master's degree in Power Engineering and Engineering Thermophysics from the same institution in 2006. In 2012, he earned his Ph.D. from the School of Architecture at the Georgia Institute of Technology in the United States.

From May 2012 to September 2013, he conducted postdoctoral research at the City University of Hong Kong. He joined Tongji University in 2013, where he has served successively as an Assistant Professor and Associate Professor; during this period, he also held a temporary secondment position as Deputy Director of the Construction and Transportation Commission of Huangpu District, Shanghai.

His primary research interests include fault diagnosis and operational optimization of HVAC systems, diagnosis and prediction of whole-building energy consumption, and the application of transfer reinforcement learning in HVAC system optimization. He has led numerous research projects, including the Young Scientists Fund of the National Natural Science Foundation of China (Grant No. 51508394) and the Shanghai Pujiang Talent Program (Grant No. 15PJ1408100). He has published over 30 academic papers in prestigious journals such as Energy and Buildings and Applied Energy.

His accolades include the Best Paper Award at the 3rd U.S. National Conference of IBPSA (2008) and a Team Award in the Student Competition at the International Conference of IBPSA (2009). Currently, he serves as the Deputy Director of the Intelligentization Professional Committee under the HVAC Industry Technology Innovation Alliance, and as a Committee Member of the Building Commissioning and Operation & Maintenance Professional Committee under the China Association of Building Energy Efficiency.

Abstract

Reinforcement learning (RL) offers a model-free approach to HVAC cooling side control but imposes a high start cost that is prohibitive for newly commissioned systems: the agent operates near-randomly in the early training stage, wasting energy before a useful policy is learned. Transfer reinforcement learning (TRL) addresses this by initializing the target agent with knowledge from a pretrained source agent. However, prior TRL work in building control has applied a single algorithm to a single type of source-to-target mismatch, leaving a critical question unanswered: does the structural form of the agent carrier — the Q-table in Q-learning, the layer-decomposed Q-network in DQN, or the full policy network in DDPG — determine how much transferred knowledge survives when the target system differs from the source in equipment quantity, capacity, and topology. This paper presents TRL methods for Q-learning, DQN, and DDPG, each evaluated across three target systems that differ from a common source system along one configuration dimension each. A component-level gray-box cooling side model serves as the simulation environment, and an exhaustive search baseline establishes a theoretical performance ceiling. The results confirm that the detailed RL algorithm is the primary determinant of transfer effectiveness, and that topology is the critical boundary condition. Equipment quantity caused the least degradation across all three algorithms, with convergence speed improving by 46–64% after transfer. Topology produced the most divergent outcomes: DQN achieved its largest convergence improvement (69.2%), while DDPG produced negative transfer — the only case in the study where transfer yielded a worse converged result than learning from scratch. Across all cases, TRL improved convergence speed by 27–69% but final steady-state performance by only 0.77–3.37%, establishing that transfer functions as a convergence accelerator rather than a performance amplifier in cooling side control.