Data Augmentation-Based Photovoltaic Power Prediction

In recent years, as the grid-connected installed capacity of photovoltaic (PV) power generation has increased by leaps and bounds, it has assumed considerable importance in predicting PV power output. However, the power prediction of newly built small-scale PV power plants often suffers from many unexpected problems, such as data redundancy, data noise, data sample imbalance, or even missing key data. Motivated by the above facts, this paper proposes a data augmentation-based prediction framework for PV power. Firstly, the daily power distance measurement is used to analyze feature correlation and filter erroneous data. Then, the autoencoder network trained based on the random discarding mechanism is used to restore the PV power generation data. At the same time, a specific data augmentation method for the PV power curve is designed to eliminate the influence of data sample imbalance. In the final experimental section, compared with the latest method, this method achieved the highest MAE accuracy of 8.26% and RMSE accuracy of 10.96%, which proves the effectiveness of this method.