New AI based weighting strategy for 2m temperature and 10m wind speed forecasting over Moroccan airports using the analog ensemble method.

In front of determinism limitations, ensemble forecasting provides competitive advantage assessing uncertainty and helping weather information users in decision-making. Analog ensemble method (AnEn) is one of the most intuitive and computationally cheap ensemble methods that leverages a single deterministic model integration to produce probabilistic information. This method builds an ensemble forecast from a set of past observations of the target variable, neatly selected from a historical training dataset. For a given location, the most similar past forecasts to the current prediction are identified and the associated  past observations are nominated  as members of the analog ensemble forecast. However, The  AnEn forecasting quality is tightly affected by the process of skillful analogs selection in the training data which depends on predictor’s weighting among other factors. This work presents a new weighting strategy based on machine learning techniques (XGBoost, Random Forest and Linear regression) and assesses the impact of its application on the AnEn performance  for 10m wind speed  and 2m temperature forecasting over 13 Moroccan airports in the short term forecasting framework (24 hours). To achieve this, hourly forecasts from the operational mesoscale AROME model and the verifying observations covering 5 year period (2016-2020) are used.  The predictors include 2m temperature, 2m relative humidity, 10m wind speed and direction, mean sea level pressure and surface pressure,  meridonal and zonal components of 10m wind. The basic configuration of Delle Monache et al. (2013) -DM13- where all the predictor’s weights are equal to one is used here as a benchmark. The best weights are computed independently from one airport to another. Since the proposed predictor-weighting strategies can accomplish both the selection of relevant predictors as well as finding their optimal weights, and hence preserve physical meaning and correlations of the used weather variables, the AnEn performances are improved by up to 50 % for bias and by 30% for RMSE for most airports. This improvement varies as function of lead-times and seasons compared to AROME and DM13’s configuration. Results show also that AnEn performance is geographically dependent where a slight worsening is found for some airports.

 

Keywords : Analog Ensemble,  Machine Learning, Predictors Weighting Strategies, 2m Temperature, 10m Wind Speed, XGBoost, Linear Regression, Random Forest, Ensemble Forecasting.