Deep Learning-based Prediction Method For Industrial Process Variables

With the continuous improvement of the information level of industrial enterprises,a great deal of data has been accumulated in their production.Such data contains important information of the production process.Accurate prediction of industrial process variables by using these data can provide helpful guidance for the operation of industrial production process.However,the existing prediction methods are not suitable for the industrial data with high noise,high nonlinearity,high coupling and missing values.By utilizing the large amount of data accumulated in the production process,this dissertation studies deep learning-based prediction method for industrial process variables.The specific research contents are as follows.Aiming at the characteristics of high noise and strong nonlinearity of industrial data,a deep denoising kernel function-based least square support vector machine prediction model is proposed,in which a deep kernel function network is formed by stacking multiple kernel functions lay by layer to replace the shallow kernel function in the least square support vector machine.Such model can extract the deep features of the sample data to deal with its high nonlinear problem.To improve the fitting performance of the model for the data with high level noises,the denoising algorithm is incorporated into the training process of the deep kernel function network.To further improve its modeling ability,the parameters are fine-tuned by using the back-propagation algorithm after pre-training each layer of the network.By using such deep denoising kernel function,the modeling ability of the least square support vector machine for high noise and high nonlinear data is improved.For the long-term prediction of industrial time series,a granular computing-based deep learning prediction model is proposed.By designing a deep granulation network,the starting point and the ending point of each granule can be determined adaptively,and the basic information granules with unequal-length of the data samples can be obtained.Then,a partially overlapping sub-block basis matrix-based deep sparse coding feature decomposition method is adopted to transform the unequal-length granules into a product of a basis matrix and a coefficient matrix layer by layer.By forecasting the coefficient matrix and mapping the result from the feature space back to the original data space,the long-term prediction result is obtained.Considering the missing values in industrial time series,a prediction model for incomplete data set based on deep bidirectional echo state network is proposed.To avoid changing the essential information of sample data in the missing data imputation process,a bidirectional fusion reservoir-based deep model is presented for extracting the deep bidirectional features of the incomplete output and input data samples at past time and future time.Based on the extracted deep features,a bidirectional echo state network is constructed to complete the prediction,and the prediction accuracy of the model is improved by unsupervised and supervised fine-tuning.Aiming at the issue of the industrial correlated time series interval prediction,an attention mechanism and deep network-based interval prediction model for correlated time series is proposed,which adopts a time-based self-attention mechanism and a multi-factor-based soft-attention mechanism to obtain the importance degree of the influence of factor variables on correlated variable.Considering the temporal influence relationship of the factor variables on the correlated variable,a deep dual daughter-cell long short-term memory network is designed for prediction.To obtain the estimated value of the prediction result and its reliability index,the Bootstrap is employed for constructing the prediction interval.To validate the performance of the proposed methods,benchmark data sets,artificial data sets and industrial actual data in the energy system of a steel enterprise in China are employed for testing.The results show that the proposed methods present the higher prediction accuracies for the related prediction problems,and the prediction intervals constructed by the proposed method present excellent indices,which can satisfy the practical prediction requirements of industrial field.