Estimation of one-minute direct normal irradiance using a deep neural network for five climate zones

Due to its high cost and significance for solar thermal technologies, the calculation of direct normal irradiance has dominated the modelling of solar radiation components. There are three methods described in the literature to calculate direct normal irradiance: regression analysis, probabilistic approach, and application of machine learning models. To demonstrate that machine learning models can achieve sufficient generality for solar radiation data with good statistical results, a new methodology is proposed with three main goals: that required only one measure data as input, trained a deep neural network model using only one station referred as the reference station, and evaluate the direct normal irradiance estimated by the model with the measured values from others stations of the same climate zone. The results then were compared against the most prominent separation models in the literature. The proposed methodology shows that the model learn the radiative characteristics of the reference station and is a feasible method to estimate the component of solar radiation in others location. Therefore, the proposed approach is a viable methodology that has the ability to reach the sought-after quasi-universal model by researchers, as demonstrated by the neural network’s similar or superior performance compared to the separation models. •Estimation of direct normal irradiance with a deep neural network for five climate zones.•A novel methodology is proposed that can be used to seek the quasi-universal model.•Evaluation of the reference station for every Köppen and Geiger climate zone.•Test the neural network trained with the reference station with the stations belonging to the same climate zone.•Comparison of the estimation of the neural network results with the most relevant separation models.