An Interpretable Soft Sensor Model for Power Plant Process Based on Deep Learning

For deep learning-based soft sensors, the lack of interpretability leading to unreliable models has become one of the most important issues. In this study, we propose an interpretable soft sensor based on deep learning and attention mechanisms. The model's network structure consists of a long short-term memory (LSTM) network with attention mechanism to form a temporal feature extraction module, and a stacked graph convolutional network with attention mechanism to form a spatial feature (relationship between features) extraction module of the model. When the model training is completed, the learned optimal graph structure, along with the attention matrix, displays the interpretability of the model in both temporal and spatial dimensions. In order to verify its performance, this model was applied to predict the thermal deformation of the power plant air preheater rotor, and its good sensing performance was confirmed by comparison with other novel soft sensors. And due to the addition of the interpretable components, it has better interpretability than other existing soft sensors. In addition, since the graph structure obtained from the proposed model can only show which features have an impact on the prediction of rotor thermal deformation, it is still unclear how these features numerically affect the prediction results. Therefore, to further understand how the soft sensor model makes its decisions, the SHAP method is used to provide numerically interpretable results, thus further improving the model's interpretability.