Under the background of green manufacturing and the “dual carbon” strategy, the carbon emission management of new energy vehicles throughout their life cycle faces problems such as strong data heterogeneity, severe information fragmentation, and insufficient modeling accuracy. In order to solve the problem of integrating multiple links, multiple variables, and strong dynamic features, a joint modeling framework based on graph neural network and improved LSTM is designed to construct a full life cycle carbon emission prediction model. By introducing multi-relation graph convolution, adaptive edge weights, and residual correction mechanisms, the structured expression and dynamic modeling of the entire process of new energy vehicles from design, production to recycling are realized. The experiment is based on real industrial data from a new energy vehicle base, and performs well in multiple indicators such as MAE, RMSE, and R2. Among them, MAE is reduced to 2.61 kgCO2e, R2 is increased to 0.942, and MAPE is 5.84%. In the robustness test, the prediction stability and migration ability are significantly better than the traditional model, and the structural adaptability indicators such as PE and RR also show higher performance. This method provides a feasible path for the digital carbon emission management of the new energy vehicle industry and has a good promotion prospect.