A novel ensemble method for residential electricity demand forecasting based on a novel sample simulation strategy. (15th September 2020)

• Record Type: Journal Article
• Authors: Zhang, Guoqiang; Guo, Jifeng
• Is Part Of: Energy. Volume 207(2020)
• Publication Date: 2020-09-15
• Subjects: Electricity demand forecasting -- Improved coupled generative adversarial stacked auto-encoder (ICoGASA) -- Integrated forecast -- Self-organizing map (SOM) -- Memristor array (MA); Power resources -- Periodicals; Power (Mechanics) -- Periodicals; Energy consumption -- Periodicals; 333.7905
• DOI: 10.1016/j.energy.2020.118265
• Languages: English
• ISSNs: 0360-5442

Abstract

Abstract: This paper presents a novel ensemble method of forecasting the residential electricity demand. Firstly, the time-series of the original input variables is filtered by unscented kalman filter (UKF), and then the incremental percentages of current and previous sample points are taken as new input features of the proposed method. Secondly, an improved coupled generative adversarial stacked auto-encoder (ICoGASA) consisting of three generative adversarial networks (GAN) is developed to generate more similar errors in weather forecast and lifestyles of different residents, with less noise. All of the three GANs are composed of two deep belief networks (DBNs), which serve as generator and discriminator, respectively. The three generators of GANs are used to simulate the samples with positive error, negative error and mixed error, respectively. Then the output of the three discriminators is integrated by memristor array (MA), and the integrated output of each ICoGASA are integrated by self-organizing map (SOM). Thirdly, the input weights of SOM are optimized by MA and a new weight updated strategy (WUS). Compared with other state-of-the-art ensemble methods, the scopes of the root mean square error (RMSE) are reduced by [8.295, 16.221] %, [15.507, 28.066] %, [20.494, 36.969] %, respectively. Highlights: Filtering input data by unscented kalman filter. Sample simulation according to an improved deep learning model. Integrated strategy by self-organizing map and memristor array. Weight updated strategy based on chaotic map sequence and memristor array.