Online prediction of photovoltaic power considering concept drift

• Published in: IEEE Power & Energy Society General Meeting pp. 1 - 5
• Main Authors: Zhang, Le, Zhu, Jizhong, Cheung, Kwok, Zhou, Jialin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.07.2023
• Subjects: concept drift; Data models; Deep learning; Online PV power prediction; orthogonal weight modification; Photovoltaic systems; Prediction algorithms; Predictive models
• ISSN: 1944-9933
• DOI: 10.1109/PESGM52003.2023.10252625

Concept drift (CD) is considered to be the source of the deterioration of the accuracy of data-driven models over time. However, the CD in photovoltaic (PV) power predictions has rarely been studied. In this paper, an online PV power prediction method is proposed, which simultaneously handles the real and virtual CD of the PV power data stream. The proposed method uses a LSTM network as a predictor and consists of CD detection and model parameter update. As for CD detection, the Energy distance between the historical and new data distribution is used as the virtual drift detection criterion. The Drift Detection Method (DDM) algorithm is used to detect real drift and define drift levels. As for model parameter update, this paper uses an orthogonal weight modification (OWM) algorithm to quickly update the parameters of the LSTM and continuously learn new data features without forgetting after the drift occurs. Finally, to verify the effectiveness of the proposed method, this paper conducts tests on public datasets. The results show that the proposed method can improve the accuracy of online PV power prediction.