An Unsupervised Data-Mining and Generative-Based Multiple Missing Data Imputation Network for Energy Dataset

Missing values are ubiquitous in energy datasets, and therefore, generative-based imputation networks have attracted extensive research interest because of their strong imputation performance. However, these networks have limited accuracy when explicit class labels are unavailable and single imputation cannot fully address the uncertainty surrounding the true values of the imputed variables. This article proposes an unsupervised data-mining-based conditional generative adversarial multiple imputation network that exploits implicit categorical information and multiple imputation to improve the robustness of the final imputation results. First, a pretraining algorithm is added to develop an auxiliary classifier combined with the corresponding implicit class labels. Then, a “fuzzy-clustering-based ordering points to identify the clustering structure” algorithm is proposed to learn the implicit categorical information. Thereafter, multiple imputation is applied to an original energy dataset to improve the reliability of the final imputation results. Experimental results demonstrate the superiority of the proposed network compared to other networks.