Intelligent Decision Support for Power Grids Using Deep Learning on Small Datasets

The larger dataset possesses the greater appliance for further data analysis and using machine learning potential. However, the unique types of complex infrastructure objects exist which can be described by the features sets specific to each object only. These include electrical networks and large power electrical equipment of railway traction substations. Datasets that are acquiring from such equipment during monitoring and maintenance procedures have high dimensional by parameter variety but small samples by its length. Because of this fact, conventional machine learning approaches based on big datasets are unsuitable for intelligent decision support systems in railway power grids. This paper proposes a novel approach to deep learning problems on small datasets by the example of deep feature selection. Pretrained models and transfer learning techniques are modified in this paper to remain robust under noise and outliers. The paper presents a new architecture for intelligent decision support with deep features selection in high dimensional size-limited datasets.