A novel two-stage reinforcement learning framework for sustainable building energy management systems

With the escalating demand for sustainable energy solutions, the need for effective and eco-friendly building energy management systems is paramount. This paper aims to pioneer an approach within the framework of Sustainable Building Energy Management Systems (SBEMS) using cutting-edge artificial intelligence technology, enhancing the intelligence and efficiency of building energy management. Our study integrates a novel two-stage reinforcement learning framework, delving into the physical models of energy prediction and building energy management, illustrated through comprehensive case study based on real-world dataset. We propose an Attention-based Deep Q-Learning (AT-DQN) framework to optimize energy management, employing an attention-based Long Short-Term Memory (LSTM) method for accurate photovoltaic power prediction, facilitating better energy utilization. Subsequently, Deep Q-Learning is utilized to optimize the charging and discharging of rechargeable batteries within SBEMS. Extensive experiments on real datasets demonstrate a remarkable reduction of 12 % in the Mean Absolute Percentage Error (MAPE) for PV forecasting. Numerical analysis and comparison validate our method, resulting in a 10 % reduction in user energy costs, a 15 % surge in energy utilization, and a noteworthy 20 % enhancement in user satisfaction. Case studies indicate the promising potential of our work in intelligent building energy management, offering users a more economical, comfortable, and eco-friendly electricity experience. We also discuss three key challenges in sustainable building energy, namely storage cost, prediction cost, and regional collaborative/competitive scheduling/trading. This work contributes to the advancement of sustainable and efficient building energy management systems.