As China has been the biggest carbon dioxide emissions country in the world and taken electric power industry as a breakthrough to build carbon market, it is necessary to promote the use of carbon continuous emission monitoring system to improve the accuracy of carbon emissions accounting. Carbon emission model is built to convert CO2 flow rate and concentration to mass. The CO2 emissions per gas inflow is a relatively stable value with a random fluctuation and will be affected by gas inflow, which is called per CO2 emissions. Based on the BP neural network algorithm, we take per CO2 emissions, gas inflow, gas turbine load, steam turbine load as input layers to obtain a real value prediction interval of per CO2 emission, and conduct carbon emission abnormal data screening. In the case study, it is proved that the neural network algorithm give an efficient way to screen big CO2 emission abnormal data.