基于分层粒度对比网络的钢铁燃气调度知识获取与建模

对于钢铁燃气系统的实时有效调度是实现企业节能降耗的关键. 考虑燃气产消过程所包含的多工况特征, 提出了一种基于分层粒度对比网络的调度知识获取与建模方法. 鉴于深度对比学习对于语义信息的处理能力, 定义和描述了一系列信息粒度, 以建立能源数据的语义表示. 为初步提取多工况调度知识, 采用长短时记忆(Long and short-term memory, LSTM)网络学习具有时变特性的粒度变量特征. 在此基础上, 利用专家经验知识定性地划分对比学习样本, 建立基于粒度对比学习的知识表征网络. 为挖掘调度数据中所包含的深层次知识, 进一步提出了基于反馈机制的分层对比网络模型, 并通过网络输出层实现调度建模任务. 实验部分采用了国内某钢铁厂高炉煤气系统的实际数据进行了多组对比实验, 结果表明所提方法获得的知识表示能够有效提高燃气系统的建模精度, 帮助实现专家级别的调度表现.

Hierarchical Granular Contrastive Network-based Knowledge Acquisition and Modeling for Gas Scheduling of Steel Industry

A real-time effective scheduling for gas system of steel industry is important for achieving energy saving and consumption reduction. Considering that the gas generation and consumption processes are characterized by multiple operating conditions, a knowledge acquisition and modeling algorithm based on hierarchical granular contrastive network is proposed. In view of the capabilities of deep contrast learning for processing semantic information, a series of information granularities are defined and described to establish a semantic representation of the energy data. To extract multi-condition scheduling knowledge, a long and short-term memory (LSTM) network is used to learn the time-varying characteristics of granular variables. On top of this, a knowledge representation network based on granular contrastive learning is established, which takes advantage of the expert knowledge to partition the contrastive samples. For digging out deep-level knowledge involved in the scheduling data, a hierarchical contrastive network is further proposed by employing a closed-loop feedback mechanism, and then scheduling modeling tasks can be addressed with output layers. The practical operation data coming from the blast furnace gas system of a domestic steel plant are utilized to perform our experiments. The results show that the multi-condition knowledge representation obtained by the proposed method would help improve the modeling accuracy of the gas system and realize human-level scheduling performance.