Dual-branch deep learning architecture for enhanced hourly global horizontal irradiance forecasting

• Published in: Expert systems with applications 2024-10, Vol.252, p.124115, Article 124115
• Main Authors: Wang, Zhijie, Tang, Yugui, Zhang, Zhen
• Format: Article
• Language: English
• Subjects: Autoregressive linear model; Deep learning; Dual-branch architecture; Feature-temporal attention; Self-Attention; Solar irradiance prediction
• ISSN: 0957-4174
• DOI: 10.1016/j.eswa.2024.124115

Solar energy, a pivotal renewable resource, faces challenges posed from the inherent randomness and volatility of solar irradiance. Effective control and utilization of solar energy systems rely on the accuracy of solar irradiance prediction models. Temporal data of solar irradiance often involves a mixture of long-term patterns represented by periodicity and short-term patterns represented by nonlinearity, for which traditional approaches are difficult to balance. In this study, a novel deep learning approach employing a dual-branch architecture is proposed to forecast hourly global horizontal irradiance. The proposed model comprises two parallel branches, namely the global temporal extractor and the local temporal extractor. Two branches effectively capture complex combinations of global and local temporal patterns from multivariable input data. To enhance the effectiveness of feature extraction, the model incorporates feature-temporal attention modules and self-attention modules within the extractors, which can re-calibrate feature weights and emphasize crucial information. To strengthen the model's robustness, an autoregressive linear model is integrated in parallel, compensating for nonlinear output. The proposed model has been validated on multiple public datasets. Experimental results show the performance of the proposed model has been greatly improved, in which achieves a remarkable 41.76% improvement of forecasting accuracy over baseline models. The superiority the proposed model is also underscored through comparative analysis against state-of-the-art methods.