Data driven approach to forecast the next day aggregate production of scattered small rooftop solar photovoltaic systems without meteorological parameters

• Published in: Energy reports Vol. 8; pp. 3221 - 3233
• Main Authors: Fentis, Ayoub, Rafik, Mohamed, Bahatti, Lhoussain, Bouattane, Omar, Mestari, Mohammed
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2022; Elsevier
• Subjects: Machine learning; PV power forecasting; Renewable energy; Time series forecasting; PV power forecasting; Time series forecasting; Renewable energy; Machine learning
• ISSN: 2352-4847; 2352-4847
• DOI: 10.1016/j.egyr.2022.02.088

Photovoltaic (PV) power has became an attractive research subject, despite the variability and uncertainty of the phenomenon, PV power forecasting is considered as a promising solution for successful PV power integration, and also an important decision making information for both, grid operators and energy traders. Forecasting the aggregated PV power in different spatiotemporal scales is very relevant for grid operators, but unfortunately, in the state of the art, few works deal with this subject. In the aim to forecast directly the next day aggregate photovoltaic power of scattered small rooftop installations based only on historical production data, we propose in this paper a hybrid nonlinear autoregressive regression model, composed from a combination of empirical mode decomposition (EMD), stepwise regression (SWR) and least square support vector regression (LsSVR) techniques. To address the lack of meteorological parameters, we propose a different data preprocessing strategy, this strategy will also allow capturing the diurnal component of the PV power time series in an efficient way. An in depth analysis of the proposed approach using publicly available database was conducted. We compared the results of the proposed approach with three other models, two nonlinear autoregressive models based on LsSVR (AR-LsSVR) and FFNN (AR-FFNN), and the third one is a persistent forecaster. Three accuracy metrics were adopted to compare the different models, the nMAE, nMBE and nMRSE. In general good results were obtained in clear and partially cloudy days, but there are few days characterized by sudden change in the time series pattern that the model cant handle. But still competitive results are obtained, the hybrid model gives an nMAE = 4.12% compared to 4.94%, 4.95% and 164.98% for AR-LsSVR, AR-FFNN and the persistent forecaster respectively. The projection of the obtained results on five recently developed models shows the effectiveness of the proposed approach. •A hybrid nonlinear model without weather parameters is proposed.•Efficient and simple method to manage the PV production data is proposed.•The proposed management method helps overcoming the lack of meteorological parameters.•The proposed management method allows capturing the diurnal component effectively.•Evaluation of the approach using publicly available PV production data.