LSTM and XGBoost Models for 24-hour Ahead Forecast of PV Power from Direct Irradiation

In this work, the photo-voltaic power forecast for the next 24 hours by combining a time series forecasting model (LSTM) and a regression model (XGBoost) from direct irradiation only is performed. Several meteorological parameters such as irradiance, ambient temperature, wind speed, relative humidity, sun position, and dew point were identified as the influencing parameters of PV power variability. Thanks to the parameter extraction and selection techniques of the XGBoost model, only the direct irradiation could be kept as the input parameters. The LSTM model was used to predict the direct irradiation for the next 24 hours and the XGBoost model to estimate the future power from the predicted irradiation. These models were developed under Python 3, the exploited data was downloaded in the PVGIS database for the city of Abomey-Calavi in Benin, and the prediction was carried out on a panel of 1000W of peak power. An experimental validation was then performed by comparing the predicted irradiance values with the measured values on site. It was obtained for the LSTM model a root mean square error of 3.66 W/m2 and for the XGBoost model a root mean square error and a regression coefficient of 1.72 W and 0.992129, respectively. These results were compared with the LSTM-XGBoost performances with irradiation, temperature, sun position, and wind speed as the inputs. It was found that the use of irradiation alone as input did not as such impair the forecast performance. The proposed method was also found to be more efficient than LSTM and CNN models used alone.