Army Research Laboratory Uses BNet to Create User-Friendly Software for Communicating Weather Forecasts and Associated Uncertainty
The Situation
The uncertainty associated with weather forecasts - the “60%” in a forecaster’s “60% chance of rain” - is one of the most valuable aspects of modern meteorology. For example, a commander might not want to send a heavy tank through an area with dirt roads when there is a greater than 50% chance of heavy rain, since the road could become muddy and the tank could get stuck. A commercial airline might not want to fly a passenger plane through an area that has more than 30% chance of heavy turbulence, since the turbulences would cause unpleasant and unsafe conditions for passengers. A logistics company might want to risk sending its delivery trucks through an area with 65% chance of flooding, since ensuring its packages arrive on time is a high priority.
However, quantifying and communicating uncertainty is one of modern meteorology’s greatest challenges. Weather forecasters today rely heavily on numerical weather prediction (NWP) information in their forecast process. The NWP process involves first using raw data from weather stations spread out over a geographical area - information like location, time, date, temperature, pressure, humidity and wind speed - to create a map of conditions over the entire area between and around those stations. The meteorologist can then use a model to predict how that map of temperature, pressure and other conditions will change in the future. Those predictions are the output of the NWP process and include information about the likelihood of rain, wind, and other weather phenomena. Through the steps of the process, errors such as faulty sensor readings and imperfections in model assumptions or physics accumulate and lead to uncertainty in the forecast. Since these errors come from a number of sources across the NWP process, they are particularly difficult to quantify.
The BNet Solution
To address the problem of clarifying uncertainty in weather forecasts based on NWP information, the Army Research Laboratory at White Sands Missile Range in New Mexico sponsored the Weather Prediction Uncertainty Management And Representation (PUMAR) project. The goal of the PUMAR project was to create a user-friendly software tool that can be used to effectively communicate weather forecasts and their associated uncertainty, resulting in better and faster decision-making based on numerical weather prediction information. Artificial Intelligence Scientists at Charles River Analytics, Inc. and Meteorologists affiliated with the Air Force and Army used BNet.Builder to model uncertainty in a NWP forecast for a real weather event that occurred over New Mexico in April 2004. The BNet.Builder model incorporated both the judgments of meteorological experts and the NWP data from the Air Force Weather Agency (AFWA). By capturing the experts’ knowledge and combining it with the NWP data, the BNet model allowed Army Research Meteorologists to graphically interact with the NWP information to represent and thus understand the types and sources of uncertainty in the forecast process.
The Outcome
The Charles River Analytics’ work on the PUMAR project has demonstrated that belief networks offer unique clarity in representing and explaining uncertainty in meteorological forecasting. Randy Lefevre, a key contributor to the PUMAR project with over 20 years of meteorological experience in the United States Air Force noted: “BNet.Builder is an excellent tool to gain an intuitive understanding of the interplay between NWP variables and the overall weather forecast. During the PUMAR project we exploited BNet.Builder to help us quantify and graphically communicate the uncertainty in cloud forecasts based on the uncertainty of humidity, temperature and condensation. The application of belief networks combined with the insight gained through cognitive engineering approaches to weather forecasting will certainly enhance the meteorologist’s ability to publicly communicate important information.”
Having demonstrated the value of representing and managing uncertainty in NWP, the next phase of the project will expand the application of belief networks to other parts of the NWP process (e.g., observation uncertainties) to more accurately capture the types and sources of uncertainty in weather forecasting.
