Purpose This study aims to address the challenges in building life cycle carbon emission evaluation, particularly the dynamic prediction deviation of carbon flow and the conflict among optimization objectives. Design/methodology/approach A hybrid model combining Long Short-Term Memory (LSTM) and Non-dominated Sorting Genetic Algorithm II (NSGA-II) is proposed. LSTM is used for accurate time-series modeling of carbon flow, while NSGA-II performs multi-objective optimization. The model integrates time-series inputs, including design parameters, material types, energy consumption and operation-maintenance cycles. A multi-layer LSTM predicts carbon emission trends at each life cycle stage. Controllable variables are then used to construct objective functions covering total carbon emissions, unit cost and construction time. NSGA-II searches for optimal solutions in a non-inferior frontier space. Findings The hybrid model achieves high prediction accuracy, with MAE ranging from 0.12 to 0.15 and RMSE from 0.16 to 0.19. Optimized results show carbon emission intensity of (18.5 ± 0.7) kgCO2/m2, cost per unit area of (3412 ± 120) CNY/m2 and construction time of (189 ± 5) days, outperforming traditional methods. Under a 30% disturbance factor, the robustness index is 0.78 with only 0.26 relative deviation, indicating high engineering stability. Originality/value This study proposes a hybrid LSTM-NSGA-II framework for simulating carbon flow over the entire building lifecycle. Through the design of an input structure embedded with stage identifiers and dynamic perturbation sensitivity analysis, this approach effectively integrates prediction and optimization. Building on existing hybrid AI models, this approach specifically enhances the cross-stage nonlinear evolution of building carbon flow and the requirements for engineering stability, providing a practical technical path for sustainable building decision-making.
An et al. (Tue,) studied this question.