To address the limitations of existing unit load device (ULD) loading optimization research, including single-objective formulations, disconnect from operational control requirements, and loosely defined compartment-compatibility constraints, a three-stage progressive ULD load planning optimization method was proposed to achieve a global balance among flight safety, profitability, maneuverability, and economic efficiency. In the first stage, payload maximization was set as the objective, and an initial loading scheme with maximum payload utilization was obtained under strict safety constraints including bearing capacity and center of gravity (CG) displacement limits. In the second stage, re-optimization was performed with the pitch moment of inertia maximization as the objective, followed by normalization to unify dimensional units. In the third stage, a comprehensive trade-off was made between the normalized pitch moment of inertia and the CG deviation from the mean aerodynamic chord(MAC) percentage, while the CG envelope deviation of the entire aircraft was precisely constrained within the flight envelope to enhance safety control accuracy. To solve this model efficiently, a genetic algorithm was improved by designing an initial population generation strategy based on a compartment-ULD compatibility matrix, proposing a constraint penalty function normalization method to balance the weights of various constraints, and introducing an elitism preservation strategy along with position-constrained crossover and mutation operations to improve convergence speed and solution quality. Finally, a B747-400F aircraft case was used to validate the optimization results. Experimental results demonstrated that, across cases of varying scale, as the number of ULDs increased, the payload utilization rate was optimized to 100%, the pitch moment of inertia was improved by up to 43.21%, and the CG deviation rate was reduced to 0, with the improved genetic algorithm converging stably within 200 generations in all cases. These results validate the effectiveness and rationality of the proposed model and algorithm.