Search

Photovoltaic (PV) power generation prediction is a significant research topic in photovoltaics due to the clean and pollution-free characteristics of solar energy, which have contributed to its popularity worldwide. Photovoltaic data, as a type of time series data, exhibit strong periodicity and volatility. Researchers typically employ time–frequency signal processing methods, like empirical mode decomposition (EMD), to smooth the data during the feature engineering stage. However, improper operations at this stage could result in information leakage. Unfortunately, many existing studies on photovoltaic prediction fail to provide sufficient details on how signal processing methods are used during model training. To address this issue, this paper proposes the similarity-day extension EMD that avoids information leakage. The proposed method is validated through experiments conducted on the PV dataset of the Desert Knowledge Australia Solar Center, using mainstream models such as GRU, LSTM, CNN-LSTM, LSTN-CNN, and Bi-LSTM. The experimental results demonstrate an average improvement of 3.67% in MAE and 5.71% in RMSE when using this method, thus verifying its feasibility and effectiveness. Moreover, the proposed method can be applied to other data processing methods that may suffer from information leakage. [ABSTRACT FROM AUTHOR]