HelioClim-3 Similarity Forecast - Methodology

Last update: Jul. 2017


In brief

  • Operational since July 2015
  • Access: graphical user interface at http://www.soda-pro.com/web-services/radiation/helioclim-3-forecast and automatic (see below)
  • Time steps: 1, 5, 10, 15, 30 minutes, hourly and daily
  • Database: HelioClim-3v4 radiation values provided at d+1. 
  • Spatial coverage: Meteosat Second Generation (-66° to +66° both in latitude and longitude)
  • Data availability: Please note that the day d should be over to provide the value for d+1. If a request is launched during day which is running, the service of persistence is applied to complete day d values, and use the solar forecast similarity method to generate day d+1 values. As a consequence, as soon as a new 15 min image is collected, this service provides different values for day d+1.
  • NB: the version  5 can not be available in real time since the input data are only available at d-2.
(click on the image to magnify)


The principle of the Solar Forecast Similarity Method algorithm is to search in the HelioClim-3 past irradiance database the closest period to the current day. The closest period is defined by considering the minimum of the square distance. Once the nearest period(s) is(are) determined, the following day(s) is(are) used as a prediction for the day following the current day.
This search for the most suited day could be very time consuming, and a lot of work has been done to reduce this time by a factor up to 50. Different parameters had been adjusted, and the best compromise is given by:

  • a temporal window of 15 days before and after the date independently of the year
  • a search in the previous 4 years is considered
  • the code is searching for the 10 closest days

The days following the 10 selected days in the past are averaged to produce the prediction. With this configuration, the bias is null ; the algorithm outperforms the persistence by 20% and the error is similar to existing methods.