A digital twin-driven multi-agent deep reinforcement learning framework for synergistic resource scheduling in...

Revolutionary cultural heritage memorial sites and sports tourism destinations in Hunan Province increasingly share transport corridors, booking platforms, and event calendars, yet remain governed by sectorally siloed management bodies that overlook cross-sector demand coupling. Static admission caps and rule-based gating cannot anticipate this coupling, producing recurrent overload at heritage memorials and parallel under-utilisation at sports venues during peak holiday periods. The objective is to design and empirically validate a Digital Twin-driven Multi-Agent Deep Reinforcement Learning framework (DT-MADDPG) that delivers synergistic resource scheduling across revolutionary heritage and sports tourism nodes while preserving cultural-protection thresholds as inviolable hard constraints rather than soft reward penalties. A data-calibrated digital-twin simulation environment is built from five empirical data streams covering 12 state-designated revolutionary heritage sites and 23 sports tourism nodes across Hunan Province: nine years of visitor-flow records, carrying-capacity registers, national sporting-event calendars, hourly observations from 14 China Meteorological Administration stations, and OpenStreetMap road topology. Five cooperative MADDPG agents, aligned with the five prefecture-level culture-and-tourism coordinating clusters, jointly optimise a four-component reward that balances synergy creation, preservation enforcement through action-projection, visitor satisfaction, and operational cost. Training is implemented in PyTorch with an OpenAI Gym compatible interface, completing two million environment interactions per scenario on a single NVIDIA RTX 3090 GPU.Performance is quantified through 500 independent Monte Carlo simulation trials across peak holiday, major sporting event, and off-season scenario regimes, benchmarked against ten competing methods spanning rule-based, linear programming, NSGA-II, model predictive control, DDPG, PPO, SAC, Decision Transformer, QMIX, and MAPPO baselines.DT-MADDPG attains a 38.6% gain in the Synergy Coordination Index, a 28.3% reduction in heritage overload incidents, and a 19.6% improvement in visitor satisfaction relative to the single-agent DRL baseline in the peak holiday scenario, with all pairwise improvements statistically significant at p < 0.001 and Cohen's d values between 1.93 and 3.12 confirmed through Wilcoxon signed-rank testing.