Data-driven methodology for building energy flexibility optimization through the application of machine learning algorithms

• Data-driven methodology for building energy flexibility optimization leveraging data-driven HVAC models with ML optimization algorithms. • Benchmarking analysis between Reinforcement Learning and Genetic Algorithms for day-ahead building energy flexibility optimization use case, considering dynamic outdoor temperatures, electricity prices and flexibility requirements • Large-scale pilot validation in the Municipality of Maia (Portugal) with the local DSO E-Redes, showing an energy bill reduction of over 5%, while guaranteeing building technical restrictions. The energy transition is completely reshaping the energy system, introducing new challenges to guarantee stability and quality of power supply. In this new context, it is fundamental to develop new solutions that leverage demand flexibility to enhance the resilience of the grid. Buildings have great potential for providing flexibility services by adjusting the operation of different resources, specially Heating, Ventilation and Air Conditioning (HVAC) systems. In the literature there are several studies that apply resistance–capacitance (RC) models and traditional optimization methods such as Mixed Integer Linear Programming (MILP) or Quadratic Programming (QP) for HVAC operation optimisation. Although RC models might be adequate for many applications, data-driven HVAC models can provide several advantages in terms of replicability and scalability, especially for buildings with complex thermal dynamics. Moreover, for highly complex systems like building HVAC systems, with a large number of interacting variables and strong nonlinearities, population-based optimization methods and machine learning algorithms can offer computationally efficient alternatives compared to MILP or QP optimisation methods. Hence, this paper proposes a data-driven methodology that integrates HVAC data-driven models and machine learning (ML) algorithms to optimize building energy flexibility. As part of the study, a benchmarking analysis has been conducted between Reinforcement Learning and Genetic Algorithms. As a result of the benchmarking analysis, it is proven that Genetic Algorithms outperform Reinforcement Learning for the specific use case considered in this study. Finally, the best performing algorithm has been validated in a real-life large-scale pilot in the Municipality of Maia (Portugal) along with the local Distribution System Operator (DSO) E-Redes (part of EDP group). The obtained results have proven that the developed methodology can reduce the energy bill by over 5% leveraging the thermal energy storage capacity of the building to provide flexibility services to the DSO while ensuring comfort conditions in real-life operation. Moreover, the developed solution has been executed in standard hardware demonstrating the scalability for the provision of aggregated flexibility from multiple buildings that can be exploited by aggregators and energy community managers.