This paper assesses the skill of probabilistic forecasts of solar irradiance in the northern regions of Chile. Raw ensemble forecast are calibrated using a parametric and a novel non-parametric machine-learning-based method. As the reference approach, the ensemble model output statistics are considered. We verify the superiority of the proposed non-parametric neural-network-based ensemble correction, resulting in more than 50 % improvement in prediction performance compared to the raw forecasts.

This study uses tree-based machine learning (ML) to forecast PM_2.5 in a complex terrain region. The models show the potential to predict pollution events with several hours of anticipation, and they integrate multiple sources of information, including in situ stations, satellite data, and deterministic model outputs. The importance analysis helps understand the processes affecting air quality in the region and highlights the relevance of external sources of pollution in PM_2.5 predictability.

Consistent definition and verification of extreme events are still lacking. We propose a new generalized spatio-temporal threshold clustering method to identify extreme event episodes. We observe changes in the distribution of extreme precipitation frequency from large-scale well-connected spatial patterns to smaller-scale, more isolated rainfall clusters, possibly leading to more localized droughts and heat waves.

The North Atlantic Oscillation (NAO) has a significant impact on seasonal climate and surface weather conditions throughout Europe, North America and the North Atlantic. In this paper, we study a number of linear and nonlinear models for a station-based time series of the daily winter NAO. We find that a class of nonlinear models, including both short and long lags, excellently reproduce the characteristic statistical properties of the NAO. These models can hence be used to simulate the NAO.

A new method for weather and climate forecast model evaluation with respect to observations is proposed. Individually added values are estimated for each model, together with shared information both models provide equally on the observations. Finally, shared model information, which is not present in the observations, is calculated. The method is applied to two examples from climate and weather forecasting, showing new perspectives for model evaluation.