Non-parametric interval forecast models from fuzzy clustering of Numerical Weather Predictions

Clustering methods are proposed and evaluated as post-processing techniques that can model the uncertainty of forecasts provided by Numerical Weather Prediction (NWP) systems. These techniques try to discover relevant information about forecast uncertainty that is inherent in the performance records of the system. We investigate the application of Fuzzy C-means clustering as a powerful unsupervised learning method to discover fuzzy sets of weather forecast situations which represent different forecast uncertainty patterns. These patterns are then utilized by different distribution fitting methods to obtain statistical prediction intervals which can express the expected accuracy of the NWP system output. Three years of weather forecast records in two weather stations are used in a set of experiments to empirically study the application of the proposed approach. Skills of the probabilistic forecasts obtained by these post-processing methods are investigated by considering cross fold validation and sampling variations. Results demonstrate that the Prediction intervals generated by the proposed procedure have a higher skill compared to baseline methods.