SWAPP: Swarm precision policy optimization with dynamic action bound adjustment for energy management in smart cities

Energy storage systems are increasingly essential for aligning renewable energy generation with consumption peaks, reducing costs, and lowering carbon emissions. Managing energy systems in smart cities is increasingly challenging due to rising energy demands and the complexities of integrating renewable sources, making effective control of storage systems crucial. Rule-based controllers (RBCs) offer predefined solutions but lack adaptability. Swarm and evolutionary algorithms provide cost-effective, stable, and scalable solutions for decision-making and optimization. However, their effectiveness in data-rich scenarios remains underexplored. This study presents a framework with a decision making algorithm named SWAPP which leverages energy usage data to learn an energy management policy in urban energy systems using swarm intelligence (SI) agents. The core components of SWAPP include an unsupervised K-Means classifier for tailored decision making, a reward mechanism that promotes strategic performance assessment through delayed feedback, a two-phase policy learning approach that combines cognitive and social learning for adaptive decision refinement with an action bound adjustment mechanism to enhance precision in policy learning. A sequential batch sampling approach is introduced which makes swarm training feasible for decision problems over long horizons. The swarm-optimized policies are tested in buildings with various energy profiles, equipped with PV panels and battery storage. Performance is evaluated based on electricity cost, carbon emissions, and grid stability. SWAPP outperforms modern RL and RBCs, demonstrating its state-of-the-art performance in the standardized CityLearn building energy management environment.