Analysis of the Effect of the Time Interval Between Samples on the Solar Forecasting

Abstract

This paper analyzes the effect of the choice of the frequency between samples in the field of solar forecasting. To perform the study, the time series of solar radiation is used, in an autoregressive mode, as the only variable, to predict a single time step. Regarding the models used for the tests, the persistent model, which serves as the baseline, and neural networks are used, the most common and increasingly elaborate: Linear, MLP, CNN1D, and LSTM models. To compare the prediction accuracy two error metrics are used: RMSE and MAE. From the results it can be deduced that the analysis of the time interval between samples is a key factor, since a bad choice can result that persistent model being as good as the best predictions of the CNN1D and LSTM models. In addition, it is shown that as the time interval between samples increases, the choice of a model and its input window becomes more important. This paper intends to serve as a first guide that allows selecting parameters to implement predictive models for solar forecasting in an existing infrastructure.