【摘要】 提出一种基于Transformer编码器和卷积神经网络相融合的数据驱动方法，建立循环流化床锅炉给煤量软测量模型，解决山西某工业锅炉缺少给煤量实时监测问题。以循环流化床锅炉运行历史数作为据模型输入，首先使用Transformer编码器将输入数据映射至高维空间，通过注意力机制提取运行数据的相关特征信息；然后使用卷积神经网络对所提取的高维特征进行低维嵌入表征；最后通过全连接层实现循环流化床锅炉给煤实时预测。在实验部分通过训练数据，不仅比较了Transformer编码器与卷积神经网络融合后及单独使用的特征提取性能，还讨论了模型各项超参数对预测性能的影响。结果显示，本融合方法的决定系数（R2）达0.969 6，表明Transformer编码器与卷积神经网络相融合，可有效预测锅炉给煤量。更多还原

【Abstract】 A novel data-driven soft sensor model was proposed that integrates Transformer encoder and convolutional neural network(CNN) to predict coal feeding rate in a CFB boiler because of the lack of real-time monitoring of coal feeding in an industrial boiler of Shanxi. The CFB boiler operating data was utilized as the model input. Firstly, the operating data was mapped into the high-dimensional space by Transformer encoder, where attention technique was employed to extract relevant features from the data. Secondly, the extracted high-dimensional features were embedded into a low-dimensional space by CNN. Finally, a fully connected layer was utilized to predict the coal feeding rate. The proposed model not only compared the feature extraction performance with Transformer encoder and CNN, but also analyzed the effect of the model hyperparameters on the prediction performance. The results showed that the determination coefficient(R2) of this integrated method was 0.969 6, demonstrating that the integration of Transformer encoder and CNN could effectively predict the coal feeding rate.更多还原