The Met Office Makes Leaps in Last Mile Forecast Generation with Amazon Nova

Overview

The UK's Met Office worked with AWS to prototype automated maritime Shipping Forecast text generation using Amazon Nova Pro and a fine-tuned Nova Lite vision-language model, achieving 62% LLM accuracy for complete forecast generation and 83% gale warning accuracy versus forecaster observations, built in four weeks.

The challenge

For key marine warning services such as the Shipping Forecast, expert meteorologists at the Met Office spend approximately 2,920 hours a year manually evaluating vast datasets spanning wind, waves, visibility, and other weather variables across 31 sea areas, then condensing that data into a prescribed text format for broadcast; the Met Office wanted to explore whether AI could augment and automate this textual weather forecast generation as data volumes grow.

The solution

Working with the AWS Specialist Prototyping Team over four weeks, the Met Office developed a discovery pathfinder prototype using Amazon Nova Foundation Models to turn raw gridded weather data into readable forecast text. It used Nova Pro in Amazon Bedrock with maritime-specific prompt engineering to convert gridded data into text descriptions and forecasts (an LLM approach via an intermediate text representation), and fine-tuned Nova Lite using Amazon SageMaker distributed training as a vision language model to process hourly weather evolution through visual pattern recognition, resulting in a single simplified workflow that could be customized to other use cases.

Reported business value

Evaluated with strict word-level comparisons against human-written forecasts over a three-month period, the prototype demonstrated 62 percent LLM accuracy for complete forecast generation, 52 percent VLM accuracy in data-to-text conversion, and 83 percent gale warning accuracy compared with forecasters' observations. The Met Office describes this as an early, significant milestone rather than an operational deployment; further refinement and testing is required before any potential AI-based improvements to the Shipping Forecast can be made operational, but the pathfinder provides a foundation for possible efficiencies across the almost 300 Met Office products and services that involve transforming raw grid data to text.