Beschreibung:
To shave peaks and fill valleys, the thermal power units are required to increase their operating time in the low-load range which negatively affects the denitrification of selective catalytic reduction (SCR). To quantify the effect, it is necessary to make an accurate prediction on the selective catalytic reduction inlet gas temperature (SCRIGT). First, the extreme gradient boosting (XGBoost) model is adopted using the unit operating parameters as inputs and load∙SCRIGT -1 as output. The prediction result shows the maximum absolute errors of 23.78% and 20.79% from two thermal power units with different type of boiler. Then, local interpretable model agnostic explanations (LIME) is applied to explain the prediction result from the XGBoost model thereby finding a linear relationship between load and load∙SCRIGT -1 with R-squared 0.994, based on which an interpretable relationship model is constructed. Finally, comparison analysis shows that the mean absolute percentage errors from the interpretable relationship model are improved from 3.07% and 2.49% to 0.68% and 0.97%, and the maximum absolute error are decreased by 85.11% and 85.98%, making the interpretable relationship model more practical to detect the SCR operation status under deep peak load regulation