Accurate solar radiation estimation is crucial for the optimal design of solar energy systems used in numerous applications. Thus

this research aims to investigate the forecasting of hourly global horizontal irradiance using both univariate and multivariate methods. Deep learning techniques

including long–short-term memory

convolutional neural networks

and a hybrid of convolutional neural networks/long–short-term memory are employed. The effects of fixed and varying learning rates are explored under the condition of a fixed window size of 48 hours. Data collected from three major cities in the United States are employed to cover a broad range of annually received solar radiation. The data are divided into three subsets: 60% are used for training

20% for cross-validation

and 20% for testing. The results revealed that the convolutional neural networks and long–short-term memory models outperform the hybrid convolutional neural networks/long–short-term memory model based on the lower values of the root-mean-squared error (RMSE)

mean absolute error (MAE)

and higher coefficient of determination (R2). For instance

the multivariate long–short-term memory with fixed learning rate (RMSE = 0.345

MAE = 0.387

R2 = 0.994) is the best-performing model for Rochester

NY

the multivariate convolutional neural networks with fixed learning rate (RMSE = 32.89

MAE = 15.35

R2 = 0.928) is the best-performing model for Seattle

WA

and the univariate convolutional neural networks with variable learning rate (RMSE = 048.2

MAE = 23.66

R2 = 0.959) is the best-performing model for Tucson

AZ. Different learning rates were shown to not significantly influence the prediction of sunlight. Furthermore

it was concluded that changing the window size does not necessarily improve performance. This study demonstrates the efficacy of variable learning rates and hybrid models in improving global horizontal irradiance forecast accuracy.