Skip to main content
Back to timeline
发表出处待核验Source publication:

Typhoon warnings stop stalling between agencies: MAESTRO runs inside a Zhejiang government command platform and cuts the warning cycle from hours to minutes

Synopsis

Zhejiang's government emergency command platform has put MAESTRO, a large-language-model-driven multi-agent system, into routine typhoon response, so observation, hazard forecasting, impact assessment and agency-facing warning drafts happen in one cross-agency workflow. Previously these steps sat in separate meteorological, hydrological and natural-resource institutions, with incompatible semantics and serial approvals, so even accurate forecasts struggled to become timely action. Across 100 typhoon scenarios MAESTRO matched expert alert levels in 98.4% of cases at a mean decision latency of 37.9 s versus 250.0 s for the expert team, and a 72-hour replay of Typhoon Lekima (2019) showed earlier location-specific warnings would have added eight hours of lead time.

Source-provided article image: MAESTRO: large language model-enabled cross-agency coordination for natural disaster early warning | npj Artificial Intelligence
Fig. 1 · Nature Machine Intelligence

Interpretation

The cross-agency coordination step of typhoon early warning can now be carried by a role-structured agent system that outputs warning recommendations for emergency-management agencies rather than issuing anything to the public. Observation, hazard forecasting, impact prediction and protective action used to sit in different institutions, with incompatible data formats and terminology, serial approvals, and alert levels set by fixed hazard thresholds that lag fast-evolving compound risks. Across 100 typhoon scenarios over Zhejiang, 10 historical events from 2016 to 2022 and 90 high-resolution WRF simulations, the multi-agent with-tool configuration agreed with ex-post official alert levels in 98.4% ± 3.4% of cases, above single-agent LLM baselines of 72.3% to 90.9%.

The system splits institutional roles into three agent types: a Situation Awareness Agent continuously reads radar, satellite, gauge and field reports and triggers workflows, an Executive Agent decomposes and routes tasks, and meteorology, hydrology, natural-resource and emergency-management specialists call their own tools to produce analyses and recommendations. Agents exchange results through a shared natural-language context that reconciles agency-specific terminology, and scoped memory keeps event context so each reasoning step stays traceable. In Zhejiang the system entered the provincial AI·1833 R disaster assistant in mid-2024, linking eight provincial departments, 11 prefectural offices and 90 county-level emergency-management units; across four typhoon responses in 2024–2025 it ran more than 80 warning cycles, and operator-verified risk assessments agreed with expert judgment about 95% of the time.

In real response settings the system produced action plans in 3 to 5 minutes that previously took 9 to 12 experts about 3 hours, while remaining deployment-ready with only minor edits. In the 72-hour full-cycle replay of Typhoon Lekima, the observation-to-warning chain took about 660 s, roughly one-ninth of the institutional baseline of about 5940 s. The same replay indicates earlier location-specific warnings and pre-positioned deployments would have let about 180,000 residents evacuate with eight extra hours of lead time; in Ningbo dispatches were scheduled about 12 hours earlier than manual judgment, and cross-prefecture reserves were diverted before road cutoffs.

Perspective

The demonstrated scope is the middle of the early-warning chain for typhoons in Zhejiang, from observation through hazard and impact assessment to agency-facing warning drafts, while public authorization and dissemination remain with the emergency operations center; the system supported multiple typhoon responses in 2024–2025 inside the provincial AI·1833 R platform.

The evaluation emphasizes decision latency, added lead time and evacuees moved, and does not yet include prospectively defined impact metrics such as injuries or fatalities averted, avoided service downtime and economic losses; the current implementation centers on meteorology, hydrology and core lifelines such as power, water and transport, leaving healthcare, telecommunications and municipal services uncovered; a full cost-benefit analysis remains to be done.