Leveraging Artificial Intelligence to improve Local Ensemble Forecasting of Reduced Visibility Conditions over Morocco using Ensemble Analog Method

Résumé

The forecasting of surface weather parameters is a key stone and the elementary input to every weather and climate application including the diagnostic of low visibility conditions (LVC). LVC are a common cause of air traffic, road, and sailing fatalities. Forecasting those conditions is an arduous challenge for weather forecasters all over the world. Surface weather parameters and visibility forecasting is mainly based on deterministic approaches, where several uncertainty sources are faced. These limitations enhanced the reflection toward using ensemble forecasting methods, which provides helpful tools for decision making. In this thesis, a new decision support system is developed based on an analog ensemble (AnEn) low computational method to predict LVC and surface weather parameters over the main airports of Morocco. An analog for a given station and forecast lead-time is a past prediction, from the same model that has similar values for selected predictors of the current model forecast at the nearest grid point. Best analogs verifying observations forms AnEn ensemble members. For this goal, we use hourly observations and forecasts from the operational meso-scale numerical model AROME, covering the 4-year period (2016-2019). This period was splitted into training (2016-2018) and testing (2019) periods. In the basic AnEn version, the selection of analogs commonly considers a small set of predictors and associates them the same weight. Herein, for surface weather parameters, we propose two novelties: First, a new weighting strategy for predictors where we use three machine learning algorithms (Linear Regression, XGBoost and Random Forest) to assign predictors weights. The stepwise forward selection method has been used exclusively for LVC forecasting. The new weighting approaches are expected to endeavor the selection of informative predictors as well as finding their optimal weights, and hence preserve physical meaning and correlations of the used weather variables. Secondly, since AnEn requires a larger training dataset to enhance the chances of finding the best analogs, we extended the search space by integrating neighboring grid points. Thus, the analog detection is based here on 16 nearest grid points. Moreover, to picture seasonal dependency, two configurations were set : The basic one where analogs may come from any past date and the restricted one where analogs should imperatively belong to a day window around the target forecast. Results analysis shows that the developed AnEn system exhibits a good statistical consistency and it significantly improves the deterministic forecast of surface weather parameters performance temporally at each station by up to 50% for Bias and 30% for RMSE for the most of the airports. This improvement varies as a function of lead-times and seasons compared to the AROME model and to the basic AnEn configuration. From a spatial perspective, given that the new space neighboring strategy maximizes the chance to find the best analogs, clear improvements were perceived for most airports. However, performances remain geographically dependent. In some airports, where topography is heterogeneous, applying this new analog searching strategy might lead to some worsening since weather observations are rare at hectometric scale. The continuous verification analysis shows that AnEn forecasting errors are location- and lead-time dependent, the ensemble dispersion becomes higher for low-visibility cases because of their weak predictibility. The new EPS system is under-dispersive for all lead times and draws a positive bias for fog and mist events. It also displays an averaged Centered Root Mean Square Error of about 1500 m for visibilities of all dataset, 2000 m for fog and 1500 m for mist cases. For probabilistic verification analysis instead, AnEn visibility forecasts are converted to binary occurrences depending on a set of thresholds from 200 m to 6000 m by a step of 200 m. It is found that the averaged Heidke Skill Score for AnEn is 0.65 for all thresholds. However, AnEn performance generally becomes weaker for fog or mist events prediction. The founded verification scores are generally close or better than verification scores in litterature.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 3

Téléchargements : 0