Abstract:

Convolutional neural network is one of the representative algorithms of deep learning, and has made remarkable achievements in many fields such as computer vision and natural language processing in recent years, among which visual geometry group (VGG) is a widely used convolutional neural network model, and shows great potential in image recognition and classification tasks. However, the application of this algorithm in the data regression prediction scenario is not good, and there are limitations of large number of parameters and large memory consumption. To this end, a new deep VGG architecture (D-VGG) is proposed, which improves the different convolution kernel configurations of block1, block2, block3, block4 and block5 in VGG, adopts a more efficient conv5-32 layout, and adds batch normalization layers in each layer of the block. A Dropout layer is added before the fully connected layer to effectively alleviate overfitting problems and accelerate the convergence of the training process. The improved D-VGG network architecture has demonstrated excellent simulation prediction performance, solving the problem of poor regression prediction performance of convolutional neural networks. The prediction performance of the network is evaluated using the data set of an ethylene cracking furnace in a factory, and compared with other machine learning models CNN, CNN-LSTM, BP neural network, SVR, etc. The experimental results show that the prediction performance of D-VGG model is superior to other models, the test set R² reaches the highest of 0.9748, RMSE decreases by 37.5% compared with VGG16 model, and its MAE, MBE and RMSE error evaluation indexes are the smallest.

Key words: machine learning, ethylene cracker, algorithm, model, prediction, neural networks

摘要：

卷积神经网络（CNN）是深度学习的代表算法之一，近年来在计算机视觉、自然语言处理等多个领域取得了显著的成绩，其中visual geometry group（VGG）是一种应用广泛的卷积神经网络模型，并且在图像识别分类任务中展现出了极大潜力。然而，该算法在数据回归预测场景中的应用效果不佳，并且存在参数数量多、占用内存大的局限性。为此，提出了一种新的深度VGG架构（D-VGG），该架构改进了VGG中的block1、block2、block3、block4和block5的不同卷积核配置，采用更高效的conv5-32布局，同时在每层block中添加批归一化层，并在全连接层之前添加Dropout层，有效缓解了过拟合问题，并加速了训练过程的收敛，改进的D-VGG网络架构展现出优异的模拟预测性能，解决了卷积神经网络对回归预测效果不佳的问题。使用某厂乙烯裂解炉装置的数据集对该网络的预测性能进行了评估，并与其他机器学习模型CNN、CNN-长短期记忆网络（LSTM）、反向传播（BP）神经网络、支持向量回归（SVR）等比较，实验结果分析表明：D-VGG模型的预测性能均优于其他模型，测试集R²最高达到了0.9748，均方根误差（RMSE）相比于VGG16模型降低了37.5%，其平均绝对误差（MAE）、平均偏差误差（MBE）、RMSE评价指标最小。

关键词: 机器学习, 乙烯裂解炉, 算法, 模型, 预测, 神经网络