Analysis of Autoregressive Predictive Models and Artificial Neural Networks for Irradiance Estimation

Objectives: A model predictive controller was designed for a DC microgrid performed in Universidad Militar Nueva Granada at Cajica campus, which requires a 24-hours estimation of solar irradiance. Methods/Statistical Analysis: Autoregressive and neural networks based predictive models were designed and tested in order to be used, as well as an Artificial Neural Network (ANN) that was trained to be compared as an alternative solution to the problem. All models were coded and simulated on MATLAB and their performance were verified and mutually compared in order to define the best forecasting approach in the target allocation. Findings: The lack of seasons and the stochastically recorded irradiance time series, caused by sudden cloudy moments are the main characteristics of the local weather behavior. Therefore, a 5-years hourly meteorological database was used to estimate and train the ARMAX, NNF, NAR and NARX models, with the main feature of using six time and meteorological variables (air temperature, solar irradiance, atmospheric pressure, day, month and hour of measurements) to estimate a single output of hourly future irradiance. All of them were tested with statistical comparison functions such as square and absolute error criteria, Retrogression coefficient (R) and autocorrelation. Application/Improvements: The results let to define the most appropriate model to be used to generate the online data required for MPC designing to assure efficient operation of DC microgrids.