Solar irradiance time series forecasting using auto-regressive and extreme learning methods: Influence of transfer learning and clustering

Solar resource forecasting is essential for an optimal energy management in smart grids using photovoltaic (PV) production. For many sites and for short time horizons (nowcasting), approaches based on the use of time series and statistical or Artificial Intelligence methods are often preferred. These methods require a long historical time series of solar radiation not always available. A practical question often arises: what happens if we want to install a PV plant in an area not covered by a consistent historical data set? Is it possible then to “delocalize” the learning using data of another location? The objective is to compare the performances of two forecasting methods (Auto-Regressive and Extreme Learning) using alternatively for the learning process solar data measured on-site, data measured on stations with similar characteristics (from the same cluster) and data collected on any other stations. 70 Spanish meteorological stations are used (with a maximal distance between stations of 2000 km and with an altitude varying between 18 and 950 m). Using transfer Learning when no solar data are available conduces to obtain results in the same range of performances than using Direct Learning. This work shows that transfer learning (based on extreme learning forecasting) is fully acceptable (about 1 percentage point higher for classical error metric nRMSE) whatever the stations used for learning, and this deviation is reduced to 0.5 percentage point when using a station related to the same cluster. Thus, the clustering seems to be not efficient to improve the reliability of the solar irradiance forecasting.