Uncertainty modeling method of weather elements based on deep learning for robust solar energy generation of building

Jiahe Wang, Masayuki Mae, Keiichiro Taniguchi
Building and Environment, Volume 266, 2022, 112115
https://doi.org/10.1016/j.enbuild.2022.112115

Abstract
During the actual use stage, the uncertainty of weather leads to errors between the performance of building energy and the simulated values. Usually, uncertainty of weather cannot be eliminated, and difficult to quantify accurately due to complex occurrence rules. Based on the influence of weather uncertainty on solar power generation, this research propose a highly reliable and high-efficiency uncertainty modeling method based on two neural networks: (i) dual-stage attention-based recurrent neural network(DARNN): due to the use of two-stage attention mechanism, as the method to achieve importance interpretability in this study. (ii) Bayesian recurrent neural network(Bayesian RNN): regard the weight and bias value as an unknown distribution, as the method of uncertainty modeling in this study. DARNN analyzes the weight value of each weather element at each time point as the importance through a single learning, and the results show that the importance of solar radiation and relative humidity to solar power generation ranks in the top two. Based on the above results, Bayesian RNN establishes the probability distribution models of solar radiation and relative humidity at each time point of a year by learning the potential correlation between weather elements and taking the historical data of low importance weather elements as input.