application of deep neural networks for load forecasting

Recently, energy demand forecasting has emerged as a significant area of research because of its prominent impact on greenhouse gases (ghgs) emission and consequently global warming. up to now, several approaches from statistical to computational intelligent have been applied in this research filed. the problems of load demand forecasting are characterized by complex and nonlinear nature and also long-term historical dependency. the literature agrees with the fact that deep learning approach is more capable in dealing with these characteristics among existing approaches. however, the recent state-of-the-art deep network models are not robust against different exogenous variables with different historical dependency. in this study, we propose a graph framework based on parallel deepnet branches to tackle this problem. this framework consists of multi parallel branches in which different kind of networks can be incorporated. these branches are concatenated as the acyclic graph to form the final structure. the parallel recurrent branches represent the historical dependency of determinants individually and this leads to better performance in case of different historical dependency in data. to demonstrate the applicability and effectiveness of the proposed framework a case study is performed. in this case study, the performance of the proposed model is examined through a comparison study with the state-of-the-art deep network models. the comparison resulted in that the proposed framework can improve the load demand forecasting by a significant margin on average.