Ensemble nowcasting of irradiance/clouds for solar energy using novel machine learning tools - can AI beat physics?

Short Description

Starting point / motivation

The transition towards an increased use of renewable energy systems is a logical consequence giving the limited amount of fossil fuel and in respect to the need of reducing the anthropogenic climate forcing. Fossil fuels and other environmental non-friendly solutions, however, have one big advantage: their usage is easy to schedule as they are not dependent on external drivers such as current and seasonal/annual weather conditions and site specific characteristics.

Renewable energy in contrast heavily depends on the prevailing weather and climate conditions. These influences affect the power production on multiple time scales, in short and medium range production of energy. Especially in PV power production the power production reacts nearly instantaneous on changing conditions such as cloud motion.

This fact is important as fluctuations in the feed-in rates of renewable energy affect in turn the grid stability and feed-in scheduling of PV production. Given the fluctuations in PV productions the integration of it into the power grid needs accurate and reliable uncertainty estimations in the nowcasting range to secure power grid stability and reduce costs.

These predictions need to fulfil certain requirements:

  1. available in near-realtime,
  2. with a high temporal forecasting frequency,
  3. a high spatial (gridded) resolution and large spatial extent, and
  4. as accurate as possible including an uncertainty estimation.

In PV production the essential needed parameters used for the production estimation are 

  • cloud cover
  • cloud motion, and
  • global horizontal irradiance.

To provide ensemble nowcasts and cover these requirements observations (satellite data, meteorological observations, PV production) are essential for every forecasting method. Existing approaches are often based on NWP information which is often outdated when the event takes place or combine only a few possible data sources.

Given the heterogeneous, non-conformal kind of time-series data forecasting methods need a new and fast approach.

Contents and goals

In the proposed exploratory project we aim at investigating the applicability of novel machine learning methods in high spatio-temporal (sub-km, sub-hourly) nowcasting of the essential parameter and if they are able to outperform physical and statistical-dynamical models.


We aim at implementing graph (convolutional) neural networks, neural ordinary differential equations, and nLASSO.

As data sources serve satellite data and products (Copernicus, Nowcasting SAF, CAMS, and others), meteorological observations, topographical data and co-variates, and PV production data.

Furthermore, we aim at investigating if an ensemble post-processing method can be adapted to nowcasting purposes using observations only delivers sufficient results when moving to sub-km and sub-hourly scale.

Project Partners


ZAMG-Zentralanstalt für Meteorologie und Geodynamik

Project partner

  • Department of Computer Science, Aalto University
  • VERBUND Energy4Business GmbH

Contact Address

ZAMG - Zentralanstalt für Meteorologie und Geodynamik
Mag. Dr. Irene Schicker
Hohe Warte 38
A-1190 Wien