基于深度强化学习的钢铁企业能源计划配置  

Energy Planning Configuration of Steel Enterprises Based on Deep Reinforcement Learning

作　　者：胡宏涛 吴定会[1] 陆申鑫 范俊岩 HU Hongtao;WU Dinghui;LU Shenxin;FAN Junyan(School of Internet of Things Engineering,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122

出　　处：《控制工程》2024年第9期1552-1560,共9页Control Engineering of China

基　　金：国家重点研发计划资助项目(2020YFB1711102)。

摘　　要：钢铁生产具有流程长、工序繁杂、工况不确定、能源介质多等生产特性,致使面向全流程的能源介质计划配置制定难,因此提出了具有自适应学习能力的深度确定性策略梯度(deep deterministic policy gradient,DDPG)算法。首先,根据生产系统的能源平衡机理构建能源平衡模型;然后,以能源转换系统最小运行能源成本为目标函数,考虑工序上下游约束和能源供需约束等模型约束,根据能源平衡模型和目标函数搭建深度强化学习下的能源优化框架,求解过程引入方差收敛的高斯噪声和动作超限惩罚函数,增大DDPG智能体的搜索范围,并提高收敛速度,使DDPG智能体在多工况下自适应调整,更新得到最优策略;最后,针对实际案例的不同工况进行仿真实验,验证深度强化学习方法在能源计划配置方面的有效性。Iron and steel production has the production characteristics of long process,complex process,uncertain working conditions,and multi-media,which makes it difficult to formulate energy media planning for the whole process.A deep deterministic policy gradient algorithm(DDPG)with adaptive learning ability is proposed.Firstly,according to the production system energy balance mechanism building energy balance model,and then the minimum energy conversion system running energy cost as objective function,considering the process upstream and downstream constraint and energy supply and demand model,such as constraints,based on the energy balance model and objective function structures,energy optimization framework under depth of intensive study,the solving process is introduced into convergence of Gaussian noise variance and overrun penalty function.The search range and convergence speed of DDPG agent are increased,so that DDPG agent can adjust itself in multiple working conditions and update the optimal strategy.Finally,the effectiveness of the deep reinforcement learning method in energy planning and allocation is verified by simulation experiments under different working conditions of real cases.

关 键 词：钢铁企业 能源计划 多工况 深度强化学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]