On October 28, the China Meteorological Administration (CMA) released its first artificial intelligence (AI) meteorological service system, "Fenghe". This system is set to promote the intelligent transformation of meteorological services.

The logo of “Fenghe”

Currently, "Fenghe" comprises 5 core modules: the Meteorological Knowledge Center, Model Square, Meteorological AI Toolbox, Intelligent Agent Factory, and Evaluation Laboratory. It is capable of understanding meteorological service requirements, generating meteorological service content, making meteorological inferences and decisions, and utilizing meteorological tools.

The interface of Meteorological AI Toolbox of “Fenghe”

How does the "intelligence" manifest?

By accessing the "Fenghe" mini-program or website version and entering weather or service-related queries, users can automatically receive answers. "Fenghe" can provide tailored, intelligently analyzed suggestions across various scenarios – from public science Q&A, weather inquiries, and risk warning advice, to fields closely linked with weather such as transportation, tourism, healthcare, logistics, and energy.

For instance, in a road trip scenario, "Fenghe" can act as a travel advisor. For a family planning a self-drive trip from Shanghai to Baiyangdian, Hebei, it might automatically suggest avoiding highway sections prone to waterlogging during thunderstorms and recommend alternative indoor attractions, facilitating the smooth communication throughout the process.

What is “1+1+N” Technology Framework?

In terms of design philosophy, "Fenghe" employs a unique “1+1+N” Technology Framework- Dual-Core, Multi-Scenario.

The first “Core” is a base model that deeply integrates general-purpose AI models with professional meteorological knowledge. Through incremental pre-training on 50 million tokens (1 token ≈ 1.5 Chinese characters) specific to the meteorological field, it achieves an accurate understanding of meteorological standards and disaster mechanisms.

The second “Core” is an intelligent agent development platform that integrates data interfaces, toolchains, and scenario-based services, enabling the model to  acess real-time data from systems like AI-abled weather models Al Nowcasting Model (CMA-AIM-Nowcast-Fenglei) and Al Global Weather Forecast Model (CMA-AIM-GFS-Fengqing).

“Multi-Scenario” involves developing specialized application-specific intelligent agents with dedicated reasoning capabilities for various users, including the public and specific industries.

What lies behind the release of “Fenghe”?

From establishing a large-scale, hundred-billion-level meteorological service corpus to utilizing Low-Rank Adaptation (LoRA), a technique for fine-tuning large language models, for knowledge enhancement and scenario-specific tuning in meteorological services, to deep reasoning technologies based on Reinforcement Learning from Human Feedback (RLHF), and multi-agent collaboration techniques tailored for meteorological service scenarios, and so on. These advancements enable "Fenghe" to enhance its requirement comprehension and scenario adaptability, allowing it to autonomously plan and execute complex meteorological service tasks while ensuring operational stability and a positive user experience.

"Fenghe" is jointly developed by CMA Public Meteorological Service Centre and Tsinghua University.

Editor: JIANG Zhiqing