The variable operating condition of thermal power units makes the data show multi-modal characteristics

which leads to the decrease of prediction accuracy of the regression soft sensor model based on shallow network structure. An improved BP neural network(back propagation neural network

BPNN) soft sensor method is studied.Firstly

the original data features are extracted by using the strong deep learning ability of stacked sparse autoencoder(SSAE)

and then the extracted features are analyzed by BPNN. The experimental results show that

the mean square error of the SSAE+BPNN soft sensor method is 0.135 8×10~(–3) and the square correlation coefficient is 0.983 2. It is proved that its prediction accuracy and generalization ability are significantly better than those of BPNN. It is applied to the soft sensor of carbon content in fly ash of a flexible peak-shaving 660 MW ultra-supercritical generator set

and the average relative error of the prediction results is 0.91%

the overall relative error is less than ±5%

indicating the method has good engineering application value.