engleski

Sensitivity tests of Kalman FIlter algorithm for forecast post-processing

Modern weather forecasting, for a long time, is based on the Numerical Weather Prediction (NWP) models. But, although the capabilities of such models are constantly improving, the errors that they exhibit are still noteworthy. This is especially the case in complex terrain, even for the highresolution mesoscale models. Computational resources are also an important factor in operational weather forecasting. The answer to the further forecast improvement could be found in the usage of statistical post-processing methods. Some of the popular methods, that improve NWP forecast on the locations where measurements are available, are the analog-based method and the method inspired by the Kalman filter (KF). Although some of the authors have performed sensitivity tests, there has not been any paper in current literature that performs detailed sensitivity tests of ratio r for various KF methods. For that reason, we present the results of a detailed sensitivity test performed to find the optimal value of the variance ratio r for different KF-based post-processing forecasts. The KF method is applied to the point-based predictions of wind speed and wind gust. The forecasting range is 72 hours. The measurements include 61 locations across the Republic of Croatia. The wind is analyzed as both a continuous and a categorical variable. The continuous verification conducted here relies on RMSE decomposition, spectral analysis and quantile-quantile plot. The appropriate categorical verification measures are used to gain better insight, such as equitable threat score, frequency bias measure, extremal dependence index and frequency measure. Even though the usage of different r-values always comes with certain trade-offs, the proposed r-values are considered optimal since they lead to excellent results for the overall data, and the results remain satisfactory even for strong wind.

post-processing, analog method, Kalman filter

nije evidentirano