Machine Learning Approach to a Low-Cost Day-Ahead Photovoltaic Power Prediction Based on Publicly Available Weather Reports for Automated Energy Management Systems

Abstract

The fully automated and transferable predictive approach based on the long short-term memory machine learning algorithm is developed for the forecasting of the photovoltaic (PV) power output. The main challenge of this approach is using publicly available weather reports and measured PV power without any technical information about the PV system. Nevertheless, the developed model is able to predict the power output of the various PV systems for all seasons with a reasonably good accuracy. This transferability of the approach is proven by the prediction of the PV power for warm and cold periods and for two different PV systems located in Oldenburg and Munich, Germany. The mean absolute scaled error of the predictions decreases with increasing the size of training set and reaches its minimum by the training with 90 days. The PV power prediction made with the publicly available weather data is compared to the predictions made with the fee based solar radiation data. The usage of the solar radiation data leads to more accurate predictions even with small training sets. Although the model with the publicly available data needs a greater training set, it still can make reasonably good predictions. Therefore, it can be applied in forecast-based energy management systems.