Abstract. This paper presents a new method to calculate variable prediction intervals (PIs) to complement numerical weather prediction (NWP) forecasts. Direct outputs of NWP models are point, deterministic predictions that provide crisp values of meteorological attributes. They can be used in a number of applications ranging from air quality and climate modeling to estimating the impact of severe weather and forecasting of energy production. However, many applications would benefit from forecasts augmented by information about their uncertainty. The most common way of describing uncertainty is the use of prediction intervals (PIs), e.g., minima, maxima, and confidence level percentages. In this paper, we apply automatic clustering of NWP model outputs to obtain conditional PIs. In particular, the described approach relies on using forecast history as a source of information for uncertainty analysis. The historical forecasts are first grouped into clusters corresponding to distinct weather situations. Each cluster is then examined to determine its specific distribution of forecast errors. This leads to forecasts with different amounts of uncertainty, depending on the forecast context. To further improve the quality of PIs, we use several derived variables and reduce the number of features using principal component analysis (PCA). We also examine several clustering algorithms to determine their suitability for PI calculation. All presented methods are empirically evaluated using a set of experiments. To establish a sound way of evaluating prediction intervals, we develop a new, improved PI quality measure. Results show that the proposed clustering-based uncertainty analysis yields prediction intervals of high resolution and accuracy.
Keywords. Statistical prediction interval, Numerical Weather Prediction (NWP), Forecast verification, Probabilistic forecast.
Weather prediction has numerous applications in various domains. Weather forecasts are typically made and reported in the form of an expected value for the attribute of interest in a particular time and location. Numerical weather prediction (NWP) models are advanced computer simulation systems that provide expected forecasts for a number of attributes. They capture physical atmospheric processes to model atmospheric behaviour. Although the deterministic interactions of these physical simulations yield the expected values of different weather attributes with high precision, such forecasts are uncertain due to the inaccuracy of initial conditions, low spatial resolution, and various simplifying assumptions .
In many applications, it is desirable that forecasts be accompanied by the corresponding uncertainties. Information about forecast uncertainty may be as significant as the forecast itself. For instance, to predict the amount of ice likely to accrete on a power transmission line as a result of an ice storm, uncertainty analysis would allow a more...