Digitalization and Intelligentization
XU Shiming, HE Zhiqian, PENG Xianyong, SHANG Zhongbao, FAN Jingwei, WANG Junlue, QU Shuyang, LIU Yang, ZHOU Huaichun
In response to the dynamic characteristics of the boiler's high-temperature reheater wall temperature, a soft-sensor model for reheater wall temperature was proposed, which integrates a sparse self-attention mechanism (SSA), a convolutional neural network (CNN), and a bidirectional long short-term memory network (BiLSTM). First, the kernel principal component analysis (KPCA) algorithm was employed to screen and reduce the dimensionality of the original candidate variables, and the top twenty-six principal components were selected as the final inputs for the model. Secondly, leveraging the advantages of CNN in capturing local correlations and BiLSTM in learning long-term sequential dependencies, the CNN-BiLSTM framework was used to capture both short-term and long-term dependencies in the time-series data. The SSA mechanism was introduced to enhance the feature extraction and modeling capabilities of the CNN-BiLSTM model by assigning adaptive weights to different feature components. Finally, simulation experiments were conducted using historical data from an in-service 1 000 MW ultra-supercritical boiler. Results show that for high-temperature reheater wall temperature prediction, the proposed CNN-BiLSTM-SSA model achieves a root mean square error (RMSE) of 4.92 ℃, a mean absolute error (MAE) of 3.81 ℃, and a mean absolute percentage error (MAPE) of 0.624 1%. These corresponding metrics are all superior to those of the CNN, LSTM, BiLSTM, CNN-LSTM, and CNN-BiLSTM models.
