Experimental validation of an LSTM-based solar irradiance forecasting model

초록

Establishing a foundation for the application of a new predictive model in small-and medium-sized buildings, where it is difficult to install measurement equipment, is associated with several challenges. In recent years, for the predictive control of buildings, research has been conducted on the development of a solar irradiance forecasting model that is capable of predicting solar irradiance at local sites using simple weather information and global data based on the LSTM machine-learning algorithm; however, in this previous study, tests were performed using simulation data only. Therefore, in this study, the solar irradiance on a target building was measured, and the effectiveness of this previously proposed model was examined using real-world data. The model was verified and categorized into three cases based on the type of data used for the LSTM learning. Cases 1 and 2 involved the prediction of the solar irradiance on a local building based on the use of global data corresponding to five and nine regions, respectively, and Case 3 involved the prediction of solar irradiance based on the learning of local solar irradiance data only. It was observed that Case 3 showed the best performance with an RSME of 32 W/m(2), followed by Case 2, which showed a similar predictive performance, with a RMSE of 39 W/m2. These results can be used for the predictive control of buildings, and it is expected that the error will improve further as more global data becomes available for learning in future.

키워드