Research on Intelligent Disaster Prediction and Collaborative Decision-Making Method Based on Multi-Agent and Knowledge Graph

To address multi-source heterogeneous disaster monitoring data, this paper proposes an intelligent prediction and collaborative control framework integrating "conflict-aware D-S evidence fusion" and "CNN-LSTM-Attention triple representation". At the evidence layer, a joint weighting and discount reflux mechanism alleviates belief polarization; at the representation layer, spatial-temporal modeling and knowledge-guided attention are fused to achieve dynamic weighting and semantic consistency of multi-source features; at the decision-making layer, feasible region projection and CVaR constraints are introduced to enhance robustness and safety under extreme operating conditions. Experimental results demonstrate that the method achieves stable improvements in metrics such as PR-AUC and recall, exhibits superior calibration and interpretability in high-conflict scenarios, and reveals the advantageous mechanism of intelligent control.