Search

[Display omitted] • Novel multiscale variational autoencoder regressor is proposed. • The multiscale CNN module can extract multiscale information. • Establishing prediction models for ethylene and propylene production for energy structure optimization. • The prediction accuracy of the proposed model on ethylene and propylene data can reach around 95% and 96%, respectively. • The proposed model can effectively increase propylene and ethylene production and reduce carbon emissions. As modern industries increase in scale and integration, the industrial process data show more complex dynamic time variability. To solve the dynamic time-varying problem of industrial process data, a novel multiscale variational autoencoder (MSVAE) based Regressor (REG) (MSVAE-REG) is proposed for production prediction and energy saving. The encoder of the MSVAE is constructed by multiscale convolutional neural network (MSCNN) to extract the multiscale temporal dynamic features, including the overall and local trend features in the Gaussian latent space. Then, the decoder of the MSVAE recovers the input data to enhance the dynamic adaptability of multiscale features. Moreover, the REG utilizes the gated recurrent unit (GRU) to build a dynamic relationship between multiscale temporal dynamic features and the predicted output. Finally, the proposed model is verified on the propylene and ethylene production datasets. Compared with the back propagation (BP) network, the extreme learning machine (ELM), the radial basis function (RBF), the convolutional neural network (CNN) and the variational autoencoder (VAE), the MSVAE-REG achieves state-of-the-art results, with the prediction accuracy respectively reaching about 95% and 96%, providing a new judgment basis for optimizing the production process and saving energy. [ABSTRACT FROM AUTHOR]

• Novel graph attention integrating spatio-temporal features is proposed. • The fusion module can adaptively control spatio-temporal interaction features. • A propylene production prediction model for energy structure optimization is built. • The proposed model increases the propylene production by about 7.6 %. • The proposed model reduces carbon emissions by about 532.0115 tons. Propylene is essential to the national economy. However, the propylene production process data have complex spatiotemporal, nonlinear and dynamic features, and the consideration of spatiotemporal effects can predict the propylene production more accurately. Therefore, a novel graph attention network (GAT) based long short-term memory network (LSTM)(GAT-LSTM) is proposed for the production prediction and energy structure optimization of propylene production processes. The GAT extracts spatial features to obtain correlations between variables, while the improved LSTM model extracts potential temporal-dependent features. The collected spatiotemporal interaction features are fused by the spatiotemporal fusion module and the adaptive control is realized. Finally, a fully connected layer maps spatiotemporal interaction features to the target domain for obtaining the output. Experimental results show that the GAT-LSTM outperforms the Conv-LSTM, the LSTM, the GAT, the Attribute-relevant distributed variational autoencoder and the Attention-LSTM. Additionally, the GAT-LSTM can provide the optimal raw material scheme for actual propylene production processes. [ABSTRACT FROM AUTHOR]