Hourly predictions of direct normal irradiation using an innovative hybrid LSTM model for concentrating solar power projects in hyper-arid regions

Abstract

Although solar energy harnessing capacity varies considerably based on the employed solar energy technology and the meteorological conditions, accurate direct normal irradiation (DNI) prediction remains crucial for better planning and management of concentrating solar power systems. This work develops hybrid Long Short-Term Memory (LSTM) models for assessing hourly DNI using meteorological datasets that include relative humidity, air temperature, and global solar irradiation. The study proposes a unique hybrid model, combining a balance-dynamic sine–cosine (BDSCA) algorithm with an LSTM predictor. Combining optimizers and predictors, such hybrid models are rarely developed to estimate DNI, especially in smaller prediction intervals. Therefore, various commonly adopted algorithms in relevant studies have been considered references for evaluating the new hybrid algorithm. The results show that the relative errors of the proposed models do not exceed 2.07%, with a minimum correlation coefficient of 0.99. In addition, the dimensionality of inputs was reduced from four variables to the two most cost-effective variables in DNI prediction. Therefore, these suggested models are reliable for estimating DNI in the arid desert areas of Algeria and other locations with similar climatic features.

Keywords

• Solar energy
• Direct normal irradiation
• Concentrating solar power operation
• Algerian big south
• Extremal optimization
• Long Short-Term Memory