PV-Net: An innovative deep learning approach for efficient forecasting of short-term photovoltaic energy production. (20th June 2021)

• Record Type: Journal Article
• Authors: Abdel-Basset, Mohamed; Hawash, Hossam; Chakrabortty, Ripon K.; Ryan, Michael
• Is Part Of: Journal of cleaner production. Volume 303(2021)
• Publication Date: 2021-06-20
• Subjects: Photovoltaic energy -- Deep learning -- Short-term forecasting -- Convolution gated recurrent units -- Residual learning; Factory and trade waste -- Management -- Periodicals; Manufactures -- Environmental aspects -- Periodicals; Déchets industriels -- Gestion -- Périodiques; Usines -- Aspect de l'environnement -- Périodiques; 628.5
• DOI: 10.1016/j.jclepro.2021.127037
• Languages: English
• ISSNs: 0959-6526

Abstract

Abstract: Although photovoltaic (PV) energy production offers several environmental and commercial advantages, the irregular nature of PV energy can challenge the design and development of the energy management systems. Precise forecasting for PV energy production is therefore of vital importance to supply consumers to improve trust in functionality of the energy management system. Stimulated by current developments in deep learning (DL) techniques as well as the promising efficiency in energy-related applications, this study introduces a novel DL architecture, called PV-Net, for short-term forecasting of day-ahead PV energy. In PV-Net, the gates of the gated recurrent unit (GRU) are redesigned using convolutional layers (called Conv-GRU) to enable efficient extraction of positional and temporal characteristics in the PV power sequences. The Conv-GRU cells are stacked in bidirectional (Bi-dir) blocks to enable modeling temporal information in forward and backward directions. The Bi-dir block is residually connected to avoid information loss across layers and to facilitate gradient flow during training. A real-world case study from Alice Springs, Australia, is employed to evaluate and compare the performance of the proposed PV-Net against recent cutting-edge approaches. The values of four performance measures demonstrate the efficiency of the proposed PV-Net in terms of prediction precision and consistency. Highlights: A novel DL model for short-term PV energy forecasting is proposed. Spatiotemporal and residual learning is combined to improve the forecasting results. The performance is validated using a real-world case study from DKPCA PV system. Extensive analysis is performed to evaluate the impact of various lengths of input. The stability and robustness of the model are demonstrated for seasonal changes.