Day-Ahead Solar Irradiance Forecasting Model

Abstract

The article describes the development of a day-ahead solar irradiance forecasting model in an hourly resolution. A statistical model based on the method of artificial neural networks with a multilayer perceptron architecture was chosen as a forecasting technique. The ground measurements of solar radiation and the weather archive, consisting of the main meteorological quantities, were used for training the model. The selecting of optimal input data combination and hyperparameter tuning were carried out according to the criterion of the minimum error. The developed model on the basis of the multilayer perceptron consists of 5 inputs, 64 neurons in a hidden layer with 1 neuron in the output layer. In the process of predicting the hourly average solar irradiance on a tilted surface, numerical weather forecast data (total cloudiness, air temperature, relative humidity, atmospheric pressure) provided by two different weather services and the cosine of the solar incidence angle as a geometric/temporal parameter were used as the input variables.