Scholarlink

Control optimization of industrial energy supply systems is essential for the global energy transition and the reduction of energy costs for industry. Deep Reinforcement Learning (DRL) can optimize the control strategy of complex energy supply systems considering volatile external influences such as electricity prices. This work implements and evaluates the DRL application for an industrial cooling supply system from the chemical and pharmaceutical industry. The simulation model of the system for training and performance evaluation is based on datasheets and historical data. Parameter identification is used to reduce the relative model error from 19% to 10% with respect to cumulative electric energy consumption over a five-month period. The DRL based operation is benchmarked against the conventional rule-based control strategy in the simulation model. The results show that the application of DRL is suitable for identifying effective and energy efficient operating strategies. However, not all operating cost factors can be reduced equally and the trade-off between different optimization targets is challenging. Based on the implementation of this work, the performance of the DRL strategy should be improved in future research.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Reinforcement Learning for Control Strategy Optimization of an Industrial Cooling Supply System in the Chemical and Pharmaceutical Industry

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年

查看全文
  • Tobias Lademann,
  • Arthur Stobert,
  • Heiko Ranzau,
  • Jonas Klingelhöfer,
  • Manuel Scharfe,
  • Matthias Weigold

摘要

Control optimization of industrial energy supply systems is essential for the global energy transition and the reduction of energy costs for industry. Deep Reinforcement Learning (DRL) can optimize the control strategy of complex energy supply systems considering volatile external influences such as electricity prices. This work implements and evaluates the DRL application for an industrial cooling supply system from the chemical and pharmaceutical industry. The simulation model of the system for training and performance evaluation is based on datasheets and historical data. Parameter identification is used to reduce the relative model error from 19% to 10% with respect to cumulative electric energy consumption over a five-month period. The DRL based operation is benchmarked against the conventional rule-based control strategy in the simulation model. The results show that the application of DRL is suitable for identifying effective and energy efficient operating strategies. However, not all operating cost factors can be reduced equally and the trade-off between different optimization targets is challenging. Based on the implementation of this work, the performance of the DRL strategy should be improved in future research.