Exploitation of a transfer learning strategy to share DRL-based controllers in a district of buildings

This thesis explores the use of advanced control strategies for district-level building energy management systems across a diverse range of European climates. The study presents a comprehensive benchmarking of 7 control methods: Rule-Based Control (RBC), Offline Deep Reinforcement Learning (Offline DRL), Online DRL, Transfer Learning (TL), Model Predictive Control (MPC), Inverse Reinforcement Learning (IRL), and Behavioral Cloning (BC) for control of mass flow rate and temperature control in 40 target buildings located in Turin, Paris, Helsinki, and Palermo. The study is implmeneded on a co-simulation coupling environment that links a Modelica-based HVAC and EnergyPlus physical building model, and Python-based controller algorithms using Functional Mock-up Interface (FMI). All building clusters consist of four different building typologies (office, retail, restaurant, and large office) combined with battery energy storage systems (BESS), photovoltaic (PV) panels, and thermal energy storage (TES) systems. Performance indicators studied encompassed operating costs, self-sufficiency, self-consumption, thermal dynamic condition, and grid dependency. The results indicate that advanced control strategies can achieve improved performance compared to rule-based control, showing approximately 30.5% improvement in cost efficiency and up to 84.13% enhancement in renewable energy self-sufficiency, depending on climate conditions and building configuration. Model Predictive Control achieved cost reductions of up to 42% in Turin and 31% in Palermo while maintaining comparable self-sufficiency levels. Deep Reinforcement Learning enhanced self-sufficiency by up to 11.7 percent in continental climates, while Transfer Learning demonstrated improved generalization capacity, increasing self-sufficiency by 9.4% in Turin without extensive retraining. Thermal performance analysis reveals that advanced controllers can effectively manage the trade-off between comfort and energy efficiency through strategic temperature constraint violations during high electricity price periods. Battery management policies show intelligent adaptation to climate-dependent solar availability, exhibiting diverse charging and discharging characteristics across the four climate zones. Advanced controllers demonstrate climate-specific adaptations: more aggressive battery cycling in Helsinki to exploit limited solar resources, balanced approaches in Paris for weather variability, optimized temperature regulation in Turin for cost minimization, and strategic export management in Palermo to utilize abundant solar generation. However, these performance improvements come with increased computational complexity, implementation challenges, and requirements for extensive training data that must be considered for practical deployment. The study demonstrates that learning-based and model-based control strategies have potential for improving building energy performance across diverse climatic conditions, with the most significant benefits observed in challenging continental climates where traditional rule-based approaches show limitations.