Hourly solar irradiance prediction using deep BiLSTM network

Accurate measurement of solar irradiance is of great significance in many applications, such as climatology, energy and engineering. Deep learning models have achieved good results in solar irradiance prediction for a single site. However, most studies take meteorological parameters as the model inputs and the irradiance values as the model outputs. Because different regions have different climates, only considering the relationship between meteorological parameters and irradiance has limitations. This paper presents a novel scheme for forecasting irradiance. The method considers the hourly irradiance prediction model to be the superposition of two parts: a daily average irradiance prediction model and the irradiance amplitude prediction model. Two submodels were constructed by using deep bidirectional long short-term memory (BiLSTM) network. For the task of irradiance prediction for 25 stations located in the United States, which are located in five different climates, the proposed method performs best for 21 stations (84%) in terms of the root mean square error, 18 stations (72%) in terms of the mean absolute error, and 17 stations (68%) in terms of the coefficient of determination. Moreover, the method adopted in this study displays a stronger irradiance prediction ability than the traditional methods for 80% of the climates included in the experiment.