Smart baseline models for solar irradiation forecasting

  1. Urraca, R. 1
  2. Antonanzas, J. 1
  3. Alia-Martinez, M. 1
  4. Martinez-De-Pison, F.J. 1
  5. Antonanzas-Torres, F. 1
  1. 1 Universidad de La Rioja
    info

    Universidad de La Rioja

    Logroño, España

    ROR https://ror.org/0553yr311

Revista:
Energy Conversion and Management

ISSN: 0196-8904

Año de publicación: 2016

Volumen: 108

Páginas: 539-548

Tipo: Artículo

DOI: 10.1016/J.ENCONMAN.2015.11.033 SCOPUS: 2-s2.0-84949293195 WoS: WOS:000369191300051 GOOGLE SCHOLAR

Otras publicaciones en: Energy Conversion and Management

Resumen

This work presents a kind of smart baseline models for solar irradiation forecasting, as models are only fed with meteorological records and solar-computed values, easy-to-obtain inputs that facilitate their implementation worldwide. Global horizontal irradiation (GHI) is predicted for horizons of 1 h in a site of Southeast Spain. Two types of approaches are undertaken: fixed models, trained just once with a global database, and moving models, where the training database is updated based on the features of the testing sample. The approaches are implemented with two machine learning algorithms, support vector regression (SVR) and random forest (RFs), along with the classic linear regression and kNN. Besides, genetic algorithms (GAs) are used to automate the training process of fixed models, a task traditionally performed based on the experience or the researcher. Significant improvements were obtained over the basic persistence methods with both approaches. In the case of moving models, results proved that the best approach to update the calibration set was by computing the Euclidean distance in the principal components space. Results of both approaches were comparable in terms of MAE and forecast skill (s), though slightly superior predictions were obtained with the moving SVR, with a forecast skill ranging from 8% to 23% and a testing MAE ranging from 49 to 64 W/m2 for the different states of cloudiness. Anyway, both approaches are valid baselines to compare new forecasting models fed with more difficult-to-obtain features, supplementing the classic but naive persistence models. © 2015 Elsevier Ltd. All rights reserved.