Solar PV Power Forecasting with a Hybrid LSTM-AdaBoost Ensemble

This is the authors' abstract. We don't add key points for this paper.

Smart grids aim at achieving unprecedented flexibility in energy management and a resilient quality of supply. However, the inclusion of variable renewable energy (RE) generation units makes the realization of a resilient supply a challenging one. One of the solutions to the achievement of the goal is the ability to predict or forecast power generation as accurately as possible from the various RE in the grid. Despite extensive research on the subject, RE generation forecasting still remains a challenge, and research is ongoing to achieve a near-perfect and efficient prediction. Deep Neural Network (DNN) algorithms have performed efficiently in areas like speech recognition, image classification, as well as forecasting tasks, such as economic time series, but have been sparsely applied in renewable energy power forecasting. This paper proposes a hybrid long short term memory (LSTM)-Adaboost ensemble method for solar power generation forecasting. It also does a comparative study of different LSTM configurations tested on solar PV data from Germany. In particular, the work in this paper looks at how well the LSTM-AdaBoost ensemble model predicts solar power compared to machine learning methods that do not use ensembles.