A number of recent trends, such as the increased power consumption in developed and developing countries, the dangers associated with greenhouse gases, the potential shortages of fossil fuels, and the increasing availability of solar and wind energy act as motivating factors for the development of more intelligent and efficient systems both on the power provider as well as the consumer side. One of the most important prerequisites for making efficient energy management decisions is the ability to predict energy production and consumption patterns. While long-term forecasting of average consumption had been extensively used to direct investments in the energy grid, short-term predictions of energy consumption became practical only recently. Most of the existing work in this domain operates at the level of individual households. However, the availability of historical power consumption data can be an issue due to concerns such as privacy, data size or data quality. Researchers have been provided with synthetic smart home energy management systems that mimic the statistical and functional properties of the actual smart grid in order to improve their access to public system models. Through developing time series to represent different operating conditions of these synthetic systems, the potential of artificial smart home energy management system applications will be further enhanced. The work described in this dissertation extends the ability to predict and control power consumption to the level of individual devices in the home. This work is made possible by several recent developments. Internet of things technologies that connect individual devices to the internet allows the remote tracking of energy consumption and the remote control and scheduling of the devices. At the same time, progress in artificial intelligence and machine learning techniques improve the accuracy of predictions. These components often form the basis of smart home energy management systems (HEMS). One of our insights that facilitates the prediction of the energy consumption of individual devices is that the history of consumption contains important information about future consumption. Thus, we propose to use a long short-term memory (LSTM) recurrent neural network for prediction. In a second contribution, we extend this model into a sequence-to-sequence model which uses several interconnected LSTM cells on both the input and the output sides. We show that these approaches produce better predictions compared to memoryless machine learning techniques. The prediction of energy consumption delivers maximum value when it is integrated with the active component of a HEMS. We design a reinforcement learning-based technique where a Q-learning model is trained offline based on the prediction results. This system is then validated only using real data from PV power generation and load consumption. Considering the scarcity of data among the smart grid users, in our third contribution, we propose the Variational Autoencoder Generative Adversarial Network (VAE-GAN) as a smart grid data generative model capable of learning various types of data distributions, such as electrical load consumption, PV power production and electric vehicles charging load consumption, and generating plausible sample data from the same distribution without first performing any pre-training analysis on the data. Our extensive experiments have shown the accuracy of our approach in synthesizing smart home datasets. There is a high degree of resemblance between the distribution of VAE-GAN synthetic data and the distribution of real data. The next step will be to incorporate Q-learning for offline optimization of HEMS using synthetic data and to test its performance with real test data.


If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu

Graduation Date





Turgut, Damla


Doctor of Philosophy (Ph.D.)


College of Engineering and Computer Science


Computer Science

Degree Program

Computer Science


CFE0009663; DP0027610





Release Date

February 2023

Length of Campus-only Access


Access Status

Doctoral Dissertation (Open Access)