Abstract: Under the "Dual Carbon" target, accurate accounting of carbon emission intensity for each boiler unit in coal-fired power plants is a fundamental prerequisite for implementing carbon quota allocation and low-carbon operation optimization. This study takes six coal-fired boilers in the power station of a thermal power plant as the research object. Based on nine operational parameters collected from the SIS (Supervisory Information System), including boiler load, main steam temperature/pressure, feed water temperature/pressure, etc., a BP (Back-Propagation) neural network-based CO? emission prediction model for a single unit is constructed. The model adopts a three-layer feed forward network architecture, with operational parameters of each boiler as inputs and CO? emission per unit time as the output. Experimental results demonstrate that the model achieves a Mean Absolute Percentage Error (MAPE) of 0.018, a Root Mean Square Error (RMSE) of 0.82, and a coefficient of determination (R2) of 0.903. This research provides an effective computational tool for thermal power plants to rapidly and cost-effectively obtain real-time carbon emission data for each unit under multi-boiler parallel operation scenarios.