Day-Ahead Forecast of Photovoltaic Power Based on a Novel Stacking Ensemble Method

Accurate prediction of photovoltaic (PV) power is the prerequisite for the safe and stable operation of the power grid with high penetration of PV. Despite various machine learning models for forecasting PV power have been developed, their accuracies are generally unstable. Toward this end, this study proposes a novel Stacking ensemble forecast model to improve the precision of day-ahead PV power forecasts. Different from the traditional Stacking model that uses the original training dataset to train the base learners, the proposed model creates multiple sub-training sets from the original training dataset to train the base learners, so as to enhance the diversity of base models and further improve the prediction accuracy. Specifically, in the proposed Stacking ensemble model, four machine learning learners, i.e., generalized regression neural network (GRNN), extreme learning machine (ELM), Elman neural network (ElmanNN), and Long shot-term memory (LSTM) neural network are incorporated, which are trained with the diverse sub-training datasets, and a variety of candidate base models are generated. For those candidate base models, the ones with the best performance are selected and integrated through a meta-model, namely the back-propagation network work (BPNN), to produce the final PV power prediction. The proposed model is evaluated using measured data from a 15kW PV power station in Ashland, Oregon, USA. Results indicate that across three weather scenarios, the performance of the novel Stacking ensemble model consistently outperforms single models and the traditional Stacking ensemble model in terms of the errors for out-of-sample forecasting, which proves the effectiveness of the developed procedure in improving PV power forecasting accuracy.