Wenninger M (2024)
Publication Language: English
Publication Type: Thesis
Publication year: 2024
URI: https://open.fau.de/handle/openfau/30693
The worldwide transformation of electricity production from fossil and nuclear energy sources to renewable energy sources is accompanied by many challenges. One of those challenges is finding an equilibrium of supply and demand – an important balance for the stability of electric grids. Production and consumption are kept in balance by adapting electricity production to consumption. Most renewable energy sources do not produce energy when demanded, but when natural conditions are suitable. As energy cannot yet be stored efficiently, over-production is as much of a problem as underproduction. Demand Response (DR) is the means for end-users to contribute to the balancing challenge. Providing the end-users with an incentive such as time-based pricing that changes according to the supply will encourage users to contribute to the equilibrium. Users’ contribution usually has implications for their daily habits and can be associated with discomfort, thus it requires a high level of involvement. Lowering the required involvement is therefore seen as an important step toward an acceptance of time-based pricing. Since the 1980s, machine learning has been seen as a solution to lower the barrier for private households to participate. The idea is to provide households with information about their electricity consumption, make recommendations on behavior changes or take automated actions. Such information can be retrieved from monitoring a household’s electricity consumption. This thesis contributes to the process of extracting information and knowledge from monitored electricity consumption in private households using machine learning. An overview of data sources and general approaches is provided. Based on this research, the Machine Learning Demand Response Model (MLDR) is introduced, defining the relation between data, knowledge, and actions. This model enhances the understanding of the individual steps required to transform monitored electricity consumption data into individual recommendations or automated actions. These steps are: data monitoring, appliance identification, appliance usage segmentation, and appliance usage prediction. For each of these steps, this thesis provides an overview of the current research state and introduces new approaches. A new monitoring system for both individual appliances and household mains is introduced. The system was used to collect a scientific dataset called Domestic Energy Demand Dataset of Individual Appliances in Germany (DEDDIAG). It contains measurements of 50 individual appliances located in 15 homes, recorded with a sample rate of 1 Hz over a period of up to 3.5 years. The dataset has been enriched with manual appliance usage annotations as well as demographic data describing the household. The system, as well as the dataset, has been published under an open-source license. Based on this dataset, an appliance category identification algorithm is introduced. The algorithm extracts features using a wavelet transformation and classifies data using the k-Nearest-Neighbor (kNN) classifier. It was evaluated and published as a challenge baseline for DEDDIAG. Next to this approach that relies on low sample rates, a high sample rate algorithm is introduced. The algorithm is based on transforming one voltage-current cycle, known as the voltage-current (V-I) trajectory, into two separate Recurrence Plots (RPs) which are then classified using a Convolutional Neural Network (CNN) in combination with Spacial Pyramid Pooling (SPP). The al-gorithm is evaluated on three different datasets and compared to previously proposed algorithms. Finding the start and stop of an appliance is the basis for deriving usage patterns. This appliance event segmentation has received little attention from other researchers, and the most commonly used algorithm, a lower-bound thresholding approach, has never been evaluated. Using the manual annotations created for DEDDIAG, this approach is evaluated using a newly introduced performance metrics called Jaccard-Time-Span-Event-Score (JTES). Together with this, a new segmentation algorithm using Support Vector Machine (SVM) is presented. Finally, based on the usage events that were determined, a combined statistical model for appliance usage prediction is introduced. It predicts future appliance usage based on the preferred time of day and the elapsed time since it was used last. It is evaluated on the GREEND dataset as well as the DEDDIAG. The thesis concludes with an outlook of potential future work.
APA:
Wenninger, M. (2024). Demand Response Management Using Machine Learning Methods (Dissertation).
MLA:
Wenninger, Marc. Demand Response Management Using Machine Learning Methods. Dissertation, 2024.
BibTeX: Download