The role of input selection and climate pre-classification on the performance of neural networks irradiance models

Neural Network (NN) models are widely accepted in the derivation of solar radiation models due to their ability to adapt to various geographic and environmental conditions. Despite this, it is unclear what training variables are required to enhance the precision of the NN model or which of them could be considered redundant; therefore, this paper intends to clarify this issue by investigating if the Köppen climate classification could be used to substitute climatic measurements. To this end, We analyzed a variety of NN architectures using 20 years of data from 1629 weather stations belonging to three different climate types (Climate A, B, and C). We found that Köppen climate sub-classification had a limited effect on the models’ performance when the information of all data types was processed together, resulting in barely noticeable improvements from 1.2% to 2.5%. However, if data were pre-classified according to climate type, the climate sub-classification input induced significant differences. Improvements up to 14% in the precision of the models were found for Climates B and A; moreover, temperature and relative humidity daily measurements could be replaced by Köppen climate information. Cross-validation analysis, using the same amount of data for all climate types, allowed us to confirm our findings for Climates A and B and revealed that data pre-classification according to climate type for Climate C, systematically increased errors from 10% to 24%, so replacing actual climatological measurements was not possible for this climate type. Revealing such patterns would facilitate the derivation of models for scenarios of limited information on temperature and relative humidity in some locations and reveals the usefulness of soft computing to go beyond understanding climate complexity.