This paper proposes a three-stage intelligent decision-making system to solve the problems of mismatched supply and demand for vocational education resources and insufficient dynamic responsiveness: Prophet-LSTM (Long Short-Term Memory) hybrid forecasting, NSGA-II (Non-Dominated Selection Genetic Algorithm II) multi-objective optimisation, and DQN (Deep Q-Network) + game theory-based dynamic scheduling. First, the Forecast-LSTM fusion model is employed to forecast the short-term and long-term demand for professional teachers, equipment and funds, and with an adaptive weight of particle swarm optimisation (PSO), the MAPE (Mean Absolute Percentage Error) is reduced by 2.4-3.2 percentage points. Secondly, an efficient-fairness-cost three-dimensional objective function is built, and NSGA-II is used to obtain the Pareto front of solution sets; the resource utilisation rate reaches 87.2% and the fair Gini coefficient is reduced to 0.19. Finally, DQN is employed for millisecond online scheduling, and Nash negotiation and Stackelberg game are introduced to reduce the conflict resolution time among government, colleges, enterprises and students by 69% and increase their average satisfaction by 9.2%. Based on the empirical results from 56 colleges and universities in a province between 2018 and 2023, it can be seen that the system still meets the requirement of a resource utilisation rate of $\ge$82% and a fairness index of $\le$0.25 in the face of sudden changes in demand, sharp reductions in resources or policy shifts; thus, the system is both stable and reasonable. The study provides an implementable AI resource allocation model for the digitalisation of vocational education.
