To solve the problem of inaccurate carbon emissions prediction during the highway construction period, a method of optimizing the back propagation(BP) neural network by particle swarm optimization (PSO) algorithm was proposed to predict carbon emissions. The 12 key indicators, including route length, subgrade length, pavement length, tunnel length, bridge and culvert length, interchange length, excavation volume, filling volume, diesel consumption, cement consumption, crushed stone consumption and steel consumption, were refined from the four dimensions of project length, construction, energy consumption and material consumption using the analytic hierarchy process (AHP). The data from 36 highway projects were used as empirical samples for model training, and a comparative analysis was conducted based on error indicators. The results show that the R² value of the obtained PSO-BP model is 0974, while the R² value of the BP model is 0890, with the former being closer to 1. Compared to the results of life cycle assessment, the PSO-BP model has a smaller deviation from the actual value than the unoptimized BP model. The four layers of the hierarchy and the selected 12 key indicators enable the prediction of carbon emissions during the design and planning stage of highway construction, providing reference for low-carbon highway construction.