Search results

(1 - 4 of 4)
NON-INTRUSIVE LOAD MONITORING AND DEMAND RESPONSE FOR RESIDENTIAL ENERGY MANAGEMENT
Title
NON-INTRUSIVE LOAD MONITORING AND DEMAND RESPONSE FOR RESIDENTIAL ENERGY MANAGEMENT
Creator
Iwayemi, Abiodun
Date
2016, 2016-05
Description
Compared to cellphone bills which itemize billing into local, international, text messaging, and data, todays electricity bills are opaque....
Show moreCompared to cellphone bills which itemize billing into local, international, text messaging, and data, todays electricity bills are opaque. Residential electricity customers receive a monthly bill detailing their aggregate energy usage, without any insight into which appliances are responsible for what proportions of their bill. We therefore created a Non-intrusive load monitoring framework that uses only data available from smart meters and the price signals from the Electric utility, and combine it with Optimal Stopping Rule-based schedulers to create a framework to equip residents with the information they need to be more energy efficient while balancing their costs and comfort. Non-intrusive load monitoring provides homeowners with detailed feedback on their electricity usage, but an open area is automated appliance labeling and the creation of generalizable appliance models that can be trained in one home, and deployed in another. Manually labeling such events to use them for disaggregating residential appliances is a costly and tedious task, and we developed two approaches for semisupervised learning of appliance signatures. The first approach uses 1-Nearest neighbor semi-supervised learning, and we developed a stopping criterion which reduces the likelihood of mislabeling appliance instances. This approach was extended to a cluster-then-label semi-supervised learning approach which can use only 3 labeled samples of each appliance to label and classify similar appliances within the home. Our approach enables the comparison of unequal length time series, and incorporates additional features extracted from the appliance time series. Finally, we develop a hybrid framework that combines detailed appliance models learned via Non-intrusive load monitoring with optimal stopping rule schedulers. We evaluated the performance of these models in terms of cost and delay, and explored the effect that errors in the real-time price and appliance models have on appliance running costs to demonstrate how our approach outperforms scheduling using only day head prices.
Ph.D. in Electrical Engineering, May 2016
Show less
APPLICATION OF MACHINE LEARNING TO ELECTRICAL DATA ANALYSIS
Title
APPLICATION OF MACHINE LEARNING TO ELECTRICAL DATA ANALYSIS
Creator
Bao, Zhen
Date
2017, 2017-05
Description
The dissertation is composed of four parts: modeling demand response capability by internet data centers processing batch computing jobs,...
Show moreThe dissertation is composed of four parts: modeling demand response capability by internet data centers processing batch computing jobs, cloud storage based power consumption management in internet data center, identifying hot socket problem in smart meters, and online event detection for non-intrusive load monitoring without knowing label. Mathematical models are constructed to fulfill the research of the four targets, and numerical examples are used to test the effectiveness of the models. The first two parts optimize jobs in Data Center in order to find the best way of utilizing the existing computing resources and storage. Mixed-integer programming (MIP) is used in the formulation. The purpose of the third part is to identify the hot socket problem in smart meter. Machine learning method has been used to locate the bad installation of smart meters by analyzing historical data from smart meters. The fourth part is non-intrusive load monitoring for residential load in houses. Signal processing and deep learning methods are used to identify the specific loads from high frequency signals.
Ph.D. in Electrical Engineering, May 2017
Show less
OPTIMAL SCHEDULING OF ELECTRIC VEHICLE'S CHARGING/DISCHARGING
Title
OPTIMAL SCHEDULING OF ELECTRIC VEHICLE'S CHARGING/DISCHARGING
Creator
Guo, Dalong
Date
2018, 2018-05
Description
The advent of Electric Vehicles (EVs) demonstrates the effort and determination of humans to protect the environment. However, as the number...
Show moreThe advent of Electric Vehicles (EVs) demonstrates the effort and determination of humans to protect the environment. However, as the number of EVs increases, charging those EVs consume large amount of energy that may cause more pressure on Grid. On the other hand, the smart grid enables two-way energy flow which gives EVs the potential to serve as distributed storage system that may help mitigate the pressure of fluctuation brought by Renewable Energy Sources (RES) and reinforce the stability of power systems. Therefore, establishing efficient management mechanism to properly schedule EV charging/discharging behavior becomes imperative. In this thesis, we consider that EVs have one charging mode, Grid-to-Vehicle (G2V), and two discharging modes, Vehicle-to- Grid (V2G) and Vehicle-to-Home (V2H). In V2G, EVs send back their surplus power to grid, while in V2H, EVs supply the power for appliances in a house. We aim to design optimal algorithms to schedule the EV’s operations. We first consider an individual residential household with a single EV, where the EV can operate at all three modes. When the EV works in G2V mode, the owner pays the cost to utility company based on the real-time price (RTP). When the EV works in V2G mode, the owner earns the reward based on the market price from utility companies. In V2H, the owner uses the EV battery to provide power to appliances in the house rather than purchasing from the utility. We propose a linear optimization algorithm to schedule the EV’s operations based on the RTP and market price subject to a set of constraints. The objective is to minimize the total cost. The results show that in general the EV chooses G2V when the RTP is low, responding to demand response. When the RTP is high, the EV tends to work as V2H to avoid buying from the utility. When the market price is high, the EVs will perform V2G to obtain more revenue. Noting that it is not practical for a single EV to perform V2G, we further consider a different scenario in which a group of EVs is aggregated and managed by an aggregator. One example is a parking lot for an enterprise. Initially only V2G is considered, that is, EVs work as energy supplies and the aggregator collects the energy from all connected EVs and then transfers the aggregated energy to the grid. Each EV needs to decide how much energy to discharge to the aggregator depending on its battery capacity, remaining energy level, and etc. To facilitate the energy collection process, we model it as a virtual energy “trading” process by using a hierarchical Stackelberg Game approach. We define the utility functions for aggregator and EVs. To start the game, the aggregator (Leader) announces a set of purchasing prices to EVs and each EV determines how much energy to sell to the aggregator by maximizing its utility based on the announced price and sends that number to the aggregator. Then the aggregator adjusts the purchasing prices by maximizing its utility based on the optimal energy values collected from the EVs and the game process repeats till it converges to an equilibrium point, where the prices and the amounts of energy become fixed values. The proposed game is an uncoordinated game. We also consider power losses during energy transmission and battery degradation caused by additional charging-discharging cycles. Simulation results show the effectiveness and robustness of our game approach. At last, we extend the game to include G2V as well for the aggregated EV group scenario. That is, EVs may charge their batteries according to the RTP so that they can sell more to the aggregator to increase the profit when the purchasing price from the aggregator is attractive. We propose a SG-DR algorithm to combine the game model for V2G and the demand response (DR) for G2V. Specifically, we adjust the utility function for EVs and then update the constraints of the game to include the DR. Subject to the duration of parking period, we solve this optimization problem using our combined SG-DR algorithm and generate EVs’ corresponding hourly charging/discharging pattern. Results show that our algorithm can increase up to 50% utility for EVs compared with the pure game model. Finally, in conclusion, we summarize our work under different scenarios. Then we analyze the potential risk and propose the future trend of EV’s development in Smart Grid.
Ph.D. in Electrical Engineering, May 2018
Show less
Data-Driven Modeling for Advancing Near-Optimal Control of Water-Cooled Chillers
Title
Data-Driven Modeling for Advancing Near-Optimal Control of Water-Cooled Chillers
Creator
Salimian Rizi, Behzad
Date
2023
Description
Hydronic heating and cooling systems are among the most common types of heating and cooling systems installed in older existing buildings,...
Show moreHydronic heating and cooling systems are among the most common types of heating and cooling systems installed in older existing buildings, especially commercial buildings. The results of this study based on the Commercial Building Energy Consumption Survey (CBECS) indicates chillers account for providing cooling in more than half of the commercial office building floorspaces in the U.S. Therefore, to address the need of improving energy efficiency of chillers systems operation, research studies developed different models to investigate different chiller sequencing approaches. Engineering-based models and empirical models are among the popular approaches for developing prediction models. Engineering-based models utilize the physical principles to calculate the thermal dynamics and energy behaviors of the systems and require detailed system information, while the empirical models deploy machine learning algorithms to develop relationships between input and output data. The empirical models compared to the engineering-based approach are more practical in a system’s energy prediction because of accessibility to required data, superiority in model implementation and prediction accuracy. Moreover, selecting near accurate chiller prediction models for the chiller sequencing needs to consider the importance of each input variable and its contribution to the overall performance of a chiller system, as well as the ease of application and computational time. Among the empirical modeling methods, ensemble learning techniques overcome the instability of the learning algorithm as well as improve prediction accuracy and identify input variable importance. Ensemble models combine multiple individual models, often called base or weak models, to produce a more accurate and robust predictive model. Random Forest (RF) and Extra Gradient Boosting (XGBoost) models are considered as ensemble models which offer built-in mechanisms for assessing feature importance. These techniques work by measuring how much each feature contributes to the overall predictive performance of the ensemble.In the first objective of this work the frequency of hydronic cooling systems in the U.S. building stock for applying potential energy efficiency measures (EEMs) on chiller plants are explored. Results show that the central chillers inside the buildings are responsible for providing cooling for more than 50% of the commercial buildings with areas greater than 9,000 m2(~100,000 ft2). In addition, hydronic cooling systems contribute to the highest Energy Use Intensity (EUI) among other systems, with EUI of 410.0 kWh/m2 (130.0 kBtu/ft2). Therefore, the results of this objective support developing accurate prediction models to assess the chiller performance parameters as an implication for chiller sequencing control strategies in older existing buildings. The second objective of the dissertation is to evaluate the performance of chiller sequencing strategy for the existing water-cooled chiller plant in a high-rise commercial building and develop highly accurate RF chiller models to investigate and determine the input variables of greatest importance to chiller power consumption predictions. The results show that the average value of mean absolute percentage error (MAPE) and root mean squared error (RMSE) for all three RF chiller models are 5.3% and 30 kW, respectively, for the validation dataset, which confirms a good agreement between measured and predicted values. On the other hand, understanding prediction uncertainty is an important task to confidently reporting smaller savings estimates for different chiller sequencing control strategies. This study aims to quantify prediction uncertainty as a percentile for selecting an appropriate confidence level for chillers models which could lead to better prediction of the peak electricity load and participate in demand response programs more efficiently. The results show that by increasing the confidence level from 80% to 90%, the upper and lower bounds of the demand charge differ from the actual value by a factor of 3.3 and 1.7 times greater, respectively. Therefore, it proves the significance of selecting appropriate confidence levels for implementation of chiller sequencing strategy and demand response programs in commercial buildings. As the third objective of this study, the accuracy of these prediction models with respect to the preprocessing, selection of data, noise analysis, effect of chiller control system performance on the recorded data were investigated. Therefore, this study attempts to investigate the impacts of different data resolution, level of noise and data smoothing methods on the chiller power consumption and chiller COP prediction based on time-series Extra Gradient Boosting (XGBoost) models. The results of applying the smoothing methods indicate that the performance of chiller COP and the chiller power consumption models have improved by 2.8% and 4.8%, respectively. Overall, this study would guide the development of data-driven chiller power consumption and chiller COP prediction models in practice.
Show less