Forecasters these days are under a lot of pressure to not only prepare myriad forecasts, but also to be aware of the weather status over large geographic areas through many observational systems, for instance, satellite, radar, automatic weather stations (AWS), lightning detectors. Failure to note the existence of a thunderstorm near an obscure airport on a day when thunderstorms are not expected can have bad results. Consequently, we have been working on an agent-based alerting system that will keep an eye on various weather phenomena and report to the forecaster when any events occur, while avoiding 'alert pollution' during interesting weather by using filtering algorithms. The pilot system we have built currently alerts on discrepancies between terminal aerodrome forecasts (TAFs) and one minute AWS pressure and temperature data for all airports over a large region of southern Australia, when wind speed exceeds various thresholds for those same airports, when thunderstorms over a similar area exceed severity criteria, and scans a volcanic ash mailing list for events pertinent to our region.

This system performs well as far as it goes, but to be truly useful, it must be future proof, ie, it must cope with a changing environment within our weather service, able to adapt to new or alternative data sources, new algorithms as they are developed, a variety of existing systems using different data structures, the dynamic nature of meteorology, and changing user interfaces. To this end, we have been exploring an extension to our agent alerting system that involves agent service description and discovery techniques. This would allow agents to automatically discover replacement source of data when existing sources become unavailable, or new sources appear that are of higher quality. It also allows agents to react to a changing environment in the short term. AS a means of grounding these ideas we are exploring a couple of scenarios: one in which aircraft flights across Australia have an associated agent instance. These transient agents then discover agents attached to individual weather events en route, and issue alerts if a change in flight planning is required. The other scenario being explored is for fire weather, in which agents are attached to wild fires and to weather events, and issue alerts if they are likely to interact.

The paper details the pilot alerting system, including the agent language involved, how the system was assembled, and the use of the system by forecasters in a trial. The paper then explores the progress being made in the broader system.