A CNN-Assisted deep echo state network using multiple Time-Scale dynamic learning reservoirs for generating Short-Term solar energy forecasting. (August 2022)

• Record Type: Journal Article
• Authors: Mustaqeem, ; Ishaq, Muhammad; Kwon, Soonil
• Is Part Of: Sustainable energy technologies and assessments. Volume 52:Part C(2022)
• Publication Date: 2022-08
• Subjects: Renewable energy -- Deep learning -- Echo state network -- Principal component analysis -- Energy forecasting; Renewable energy sources -- Periodicals; Energy development -- Technological innovations -- Periodicals; Electric power production -- Periodicals; Energy storage -- Periodicals; 333.79
• DOI: 10.1016/j.seta.2022.102275
• Languages: English
• ISSNs: 2213-1388

Abstract

Graphical abstract: Highlights: The model introduce the output feedback into the state update equation to speed up learning. The model introduce the dynamic Bayesians algorithm to optimize the reservoir parameters. Utilizing novel algorithms to determine the number of reservoirs in order to existence methods. Our model achieves a reduce error rate, MAE 0.038, MAPE 3.331, and RMSE 0.310 on solar data. The superiority of results showing the highest reliability and robustness among baseline methods. Abstract: The integration of renewable energy generation presented an important development around the globe and conveys countless financial, commercial, and environmental assistance. However, the random and non-linear nature of solar power generation, the significant perception of solar system may challenge the comprehension and operation of the existing power system. Nowadays, a smart energy forecasting system is important for solar energy generation to provide high-quality energy for end-users to enhance the power system's reliability and operation. Inspired by the performance and advancements of deep learning approaches in the field of energy, a CNN-assisted deep echo state network (CNN-DeepESN) with principal components analysis (PCA) method is proposed for generating short-term solar energy forecasting. In this study, we designed a hybrid deep learning method to utilized one-hour solar forecasting ahead with five-minute time intervals. The CNN is used to prepare to compose the initial tensor of input solar time-series data and fed to the DeepESN model to compute the features with a high dimensional space by reservoir that avoid the model complexity due to its untrained and sparse nature. Therefore, we adopt the PCA for dimensionality reduction to minimize the computational cost since it offers significant advantages. We conducted extensive experiments to demonstrate and evaluate the proposed system performance by using authentic data collected from the Australian city, Alice Springs. Additionally, conducted a comparative analysis with other state-of-the-art models to illustrate the superiority of the proposed system. Three main evaluation matrices such as MAE, MAPE, and RMSE are used and achieves 0.0381, 3.3313, and 0.3101 score, respectively to show the model robustness and significance in energy forecasting and generation.