Reinforcement learning-based energy management system in the complex electric tariff environment

The increasing integration of distributed energy resources has transformed power systems from traditional one-way structures into two-way interactive networks. In response, the deployment of energy storage systems at the distribution level is accelerating due to their rapid charge/discharge capabilities. As electricity tariff structures become increasingly complex, optimizing the operation of energy storage systems under such conditions has become a critical challenge for energy management systems. This paper proposes a two-stage battery management system designed to minimize operational costs in environments with complex tariff structure. In the first stage, short-term power demand is predicted by statistical approach, which achieves a prediction accuracy of 94.59%, outperforming deep learning techniques while maintaining a fast prediction and fitting time of less than one second. Based on these predictions, a Min–Max linear optimization method is applied to set an appropriate threshold, effectively reducing peak consumption. In the second stage, a battery management system based on the proximal policy optimization algorithm is introduced as a real-time operation technique, which continuously adjusts charging and discharging strategies in real-time to minimize overall operational costs with consideration of time-of-use price, baseline penalty, and battery degradation cost. Simulation results using real-world power consumption data from South Korea demonstrate that the proposed method reduces overall operational costs by 4.43% compared to conventional battery management system. The results validate the practicality and effectiveness of the proposed framework for real-time battery management under complex electricity tariff structures. • Novel BMS algorithm is proposed to minimize electricity costs in South Korea. • ARIMA-based prediction model is applied to achieve short-term demand prediction. • Min-Max optimization sets daily upper bound, reducing peak demand efficiently. • PPO-based reinforcement learning is proposed for dynamic real-time battery control. • RL agent adapts to prices, baseline costs, and degradation, cutting costs 4.43%.