Ontology and Bayesian decision networks for supporting the meteorological forecasting process

The weather impacts on all aspects of human endeavour. Accurate forecasts are therefore important both to individuals and to industries for which the weather has strong economic consequences. Yet the complexity of weather systems, our lack of data about past weather, and ongoing changes in climate and local meteorology make weather forecasting inherently uncertain and extraordinarily difficult. These factors underscore the importance of computerised assistance to weather forecasters. However, many existing decision support tools for weather forecasting do not deal explicitly with uncertainty. Bayesian Networks, which are based on probability theory, and their extension to Bayesian Decision Networks, which are based on utility theory, are widely accepted as intuitively appealing, practical representations of knowledge that can be used for reasoning under uncertainty. This capacity, while a prerequisite for any technology aiming to support the weather forecasting process, is not in itself sufficient to ensure successful adoption ofthe technology by weather forecasters. Analysis of current weather forecasting practice identified the following issues: the dynamic nature of the domain requires fast and flexible development and deployment of networks; weather forecasting is a huge, multifaceted task that needs to be broken down into manageable pieces, and forecasters themselves should be able to construct networks to deal with these pieces; current weather forecasting practice includes tacit, undocumented knowledge; the large quantity of information involved and ways it is documented and captured XVIII need a bridging system to enable its use in Artificial Intelligence while remaining sufficiently intuitive to be used by forecasters: i.e. a mid-layer is needed between the way forecasters think and the stringent formalism of Artificial Intelligence. These issues motivate the overall approach in this thesis, which is to give forecasters a framework through which they may develop their own small, problem-specific Bayesian Decision Networks. In this thesis we present an integrated general framework for the development of Bayesian Decision Networks by domain experts to support the weather forecasting process. This framework has two key aspects. First, it provides a comprehensive, detailed methodology for knowledge engineering with Bayesian Decision Networks, which we call KEBN-DN, to guide forecasters in the development of Bayesian Decision Networks. Second, the framework incorporates a design and prototype implementation for an extendable ontology—an explicit representation of knowledge for sharing and re-using—for capturing the domain experts' (forecasters’) knowledge, which we call WeathOntology. This allows forecasters to capture their own weather forecasting process knowledge at the ontological level rather than moving directly into Bayesian Decision Network formalism. In addition, we provide them with a template that uses the information in the ontology to generate initial BDN structures. Through case studies we demonstrate the usefulness of the overall framework: the KEBN-DN methodology, WeathOntology as a foundational representation of knowledge, and the template. We identify criteria and metrics for evaluating Bayesian Decision Networks in the weather forecasting domain and, using these criteria, present an evaluation of two case studies: Bayesian Decision Networks for location-specific fog forecasting, and a more general Bayesian Decision Network for thunderstorm tracking. This research makes contributions to the fields of weather forecasting and Bayesian Networks technology, as well as making a novel specific use of ontology in knowledge engineering.