基于样本优化的BP神经网络SCR脱硝催化剂体积设计

Design of SCR catalyst volume based on BP neural network with optimized net training

选择性催化还原(SCR)是目前主流的烟气脱硝技术,在SCR脱硝系统设计中,催化剂体积准确设计是事关脱硝效率和成本支出的重要环节。传统的SCR催化剂体积设计主要采用经验公式计算,但由于SCR催化剂体积设计的影响因素众多,使得经验公式不仅复杂且适用性较窄。基于此提出基于样本优化的BP神经网络催化剂体积设计预测方法。研究结果表明,针对已构建的BP神经网络模型,通过平均影响值算法(MIV)对原始样本进行输入参数筛选和优化,降低了神经网络复杂性,提高了模型精度,平均预测误差从15.46%下降至10.32%。进一步通过BP神经网络预测筛选辨识样本中的不良数据,基于剔除了不良样本训练的BP神经网络,平均预测误差可进一步下降至2.55%,这一方法可为催化剂体积设计提供参考和指导。

Selective catalytic reduction(SCR) of NOx is currently the most dominant flue gas denitrification technology.During the design of the SCR system, it is important to determine the catalyst volume which directly affects the denitrification efficiency and cost. Conventional SCR catalyst volume is calculated by empirical formulas, but due to the complex factors related to the catalyst volume design, the empirical equations are very complex with limited application scopes. In this paper, a new SCR catalyst volume design method has been proposed based on BP neural network with optimized net training. The results show that the optimization of input parameters with the mean impact value(MIV)method significantly reduces the complexity of the neural network and improves the prediction accuracy, with the average prediction error decreasing from 15.46% to 10.32%. Moreover, through the identification and elimination of bad data in the original examples, the prediction error of the BP neural network can be further reduced to 2.55%. These results indicate that this method can provide good guidance for the SCR catalyst volume design.