基于知识蒸馏的锅炉飞灰含碳量预测研究  

Research on Prediction of Carbon Content in Boiler Fly Ash Based on Knowledge Distillation

作　　者：路宽 杨兴森 王海超 刘科 杨子江 LU Kuan;YANG Xingsen;WANG Haichao;LIU Ke;YANG Zijiang(State Grid Shandong Electric Power Research Institute,Jinan 250003,China;College of Electrical Engineering and Automation,Shandong University of Science and Technology,Qingdao 266590,China)

机构地区：[1]国网山东省电力公司电力科学研究院,山东济南250003 [2]山东科技大学电气与自动化工程学院,山东青岛266590

出　　处：《山东电力技术》2021年第6期6-9,34,共5页Shandong Electric Power

基　　金：国家自然基金青年科学基金项目(61803233)。

摘　　要：提出一种基于知识蒸馏框架的火电厂锅炉飞灰含碳量预测模型,以在模型输入值采样周期(分钟级)和实际飞灰含碳量采样周期(小时级)存在差异时,给出准确的分钟级飞灰含碳量预测结果。首先,选取多层前馈网络作为Teacher模型,修改损失函数的计算周期来使模型的输出与实际飞灰含碳量采样周期保持一致,完成第一次监督学习。其次,采用Xgboost作为Student模型,以Teacher模型的分钟级飞灰含碳量输出作为目标,完成分钟级输入、输出的二次监督学习。最后,选取山东省内某电厂的在运机组进行试验。结果表明,知识蒸馏方法可以对该机组锅炉的飞灰含碳量给出稳定、准确的预测。A prediction model based on knowledge distillation was proposed to give accurate prediction results of carbon content in fly ash of thermal power plant boiler in minute level when the sampling period of model input value(minute level)is different from that of actual fly ash carbon content(hour level).Firstly,the multilayer feedforward network was selected as the Teacher model,and the calculation period of the loss function was modified to make the output of the model consistent with the sampling period of the actual fly ash carbon content,so as to complete the first supervised learning.Secondly,Xgboost was used as the Student model to complete the second supervised learning of minute level input and output,with the output of minute level fly ash carbon content of Teacher model as the goal.Finally,the unit in operation of a power plant in Shandong province was selected for test.The results show that the knowledge distillation method can give a stable and accurate prediction of the carbon content in fly ash of the boiler.

关 键 词：锅炉飞灰含碳量 知识蒸馏 多层前馈网络 随机森林 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]