MEDEA - Meteorologically induced extreme events detection for renewable energy using data driven methods: from weather prediction to climate time scales

The MEDEA project, funded by the Austrian Climate Research Program, deals with the identification, detection and prediction of meteorologically induced extreme events for the renewable energy sector. The overall aim of the proposed project is to define and detect extreme events and outliers in meteorological time series relevant for renewable energies and to use these findings to improve both weather and climate predictions of such extreme events with specific focus on the Austrian renewable energy system.

A complicating factor in this research is that the definition of extreme events varies between the scientific fields and sectors. The project is structured in three parts with part I defining  extreme events. An extreme event in terms of meteorology can be an extreme event for one or several renewable energy systems but it does not have to be. Thus, defining the extreme events with relevance for renewable energy systems is a crucial and non-trivial task in this project. The definition of extreme events in renewable energy is two-fold: as first step wind, solar, and hydropower extreme events will be defined separately in accordance with stakeholders. Combined extreme events, such as calms and droughts arising together with high temperatures increasing the needs of electricity for cooling, will be described too.

Data driven methods (part II) are used to identify these extreme events within the meteorological observations and analysis data. The goal of this part is to develop i) novel clustering method that integrates the heterogeneous data collected Part I in order to enable a joint prediction of extreme events. A novel method learning a joint low-dimensional vector space embedding a large amount of point and gridded observational data is implemented. Data clustering is applied to build a supervised classifier for the prediction of extreme events This will facilitate the learning of ii) Granger causal models and of supervised classifiers reducing the input data to a manageable set of spatio-temporal factors influencing the formation of rare and frequent extreme events. Furthermore, iii) a novel anomaly detection method identifying observation patters that do not fit the normal spatial-temporal observation patterns in the different clusters of the data is developed.

To forecast such events we will use machine learning methods in project part III. Different ML and post-processing methods will be adapted for heavy tails using extreme value theory and applied to improve prediction of such events. Depending on the type of events, a two-fold modeling strategy can be anticipated using an additional model suitable for the pre-detected event using information of part II of the project. Furthermore, the return periods and frequencies of extreme events relevant for renewable energy are calculated. 

Here, we present first outcomes after project year 1.