With the rapid increase in size and computational complexities of power systems, the need for powerful computational models to capture strong patterns from energy datasets is emerged. In this thesis, we provide a comprehensive review on recent advances in deep neural architectures that lead to significant improvements in classification and regression problems in the area of power engineering. Furthermore, we introduce our novel deep learning methodologies proposed for a large variety of applications in this area. First, we present the interval deep probabilistic modeling for wind speed forecasting. Incorporating the Rough Set Theory into deep neural networks, we create an accurate interval model for point prediction of intermittent wind speed datasets. Then, we develop a graph convolutional neural network for the spatiotemporal prediction of wind speed values in multiple neighboring wind sites. Our provided numerical results show the great improvement of prediction accuracy compared to classic deep learning. Using the concept of graph convolutions, we also develop a new conditional graph variational autoencoder to learn the probability density of future solar irradiance given the historical solar irradiance of multiple photovoltaic energy sites. This study led to the state-of-the-art performance in probabilistic solar prediction in power systems domain. Moreover, we introduced a novel multimodal deep recurrent structure that makes use of both system-wide power and voltage measurements as well as load parameters for accurate real-time load modeling. The numerical results show the significant improvement of this method compared to classic deep learning in estimating dynamic load parameters of smart grids. Moreover, we develop deep dictionary learning as a new paradigm in machine learning for energy disaggregation and behind-the-meter net load decomposition. The presented work leads to the best accuracy in comparison with recent sparse coding and dictionary learning-based decomposition methods in the literature. Finally, a novel deep generative model is introduced to learn the probability density of the measurements on the nodes and edges of a power grid. Using this model, we take a large number of samples from the probability distribution of the structure of power systems, hence, generating synthetic power networks with the same topological and physical behaviors as the original power system. Our simulation results on real-world datasets show the great improvements of the proposed approach compared to the data-driven approaches in the recent literature.
Electrical and Computer Engineering
Prof. Jianhui Wang
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Khodayar, Mahdi, "Learning Deep Architectures for Power Systems Operation and Analysis" (2020). Electrical Engineering Theses and Dissertations. 41.