Abstract: To use accurate control strategy to optimize the combustion control of heating furnace for solving the core problem of combustion device optimization control in metallurgical enterprises and making the accurate prediction of furnace temperature,and to study the applicability of neural network in furnace temperature prediction,a prediction method of furnace temperature based on transfer learning and knowledge distillation was proposed.The temperature prediction model of the source zone based on the temporal convolution network was established,and the model transfer was completed by using the generative adversarial loss for domain adaptation,so as to realize the accurate prediction of the temperature in all heating zones.A distillation network based on multi-task learning was further established,which solved the shortcomings of the high delay of the deep transfer network by the way of teacher assisting student.The experimental results showed that the proposed transfer learning network could significantly improve the accuracy of furnace temperature prediction,the distillation network could significantly reduce network parameters and greatly improve the timeliness of furnace temperature prediction.

Key words: heating furnace, transfer learning, temporal convolution network, knowledge distillation

摘要： 为了采用精确的控制策略对加热炉的燃烧情况进行优化控制,解决冶金企业中燃烧装置优化控制的核心问题,对加热炉内所有加热区的温度进行预测,并研究神经网络在炉温预测方面的适用性,提出基于迁移学习和知识蒸馏的炉温预测方法。建立基于时间卷积网络的源域温度预测模型,采用生成对抗损失进行域自适应来完成模型迁移,准确预测所有加热区的温度。进一步建立基于多任务学习的蒸馏网络,该网络通过教师辅助学生的方式解决深度迁移网络延时高的缺点。实验结果表明,所提迁移学习网络可以明显提升炉温预测的准确性,蒸馏网络可以明显减少网络参数,极大提高炉温预测的时效性。

关键词: 加热炉, 迁移学习, 时间卷积网络, 知识蒸馏